2504.07491

# Kimi-VL Technical Report **Authors**: Kimi Team \addbibresource template.bib Abstract We present Kimi-VL, an efficient open-source Mixture-of-Experts (MoE) vision-language model (VLM) that offers advanced multimodal reasoning, long-context understanding, and strong agent capabilities —all while activating only 2.8B parameters in its language decoder (Kimi-VL-A3B). Kimi-VL demonstrates strong performance across challenging domains: as a general-purpose VLM, Kimi-VL excels in multi-turn agent tasks (e.g., OSWorld), matching flagship models. Furthermore, it exhibits remarkable capabilities across diverse challenging vision language tasks, including college-level image and video comprehension, OCR, mathematical reasoning, multi-image understanding. In comparative evaluations, it effectively competes with cutting-edge efficient VLMs such as GPT-4o-mini, Qwen2.5-VL-7B, and Gemma-3-12B-IT, while surpassing GPT-4o in several key domains. Kimi-VL also advances in processing long contexts and perceiving clearly. With a 128K extended context window, Kimi-VL can process diverse long inputs, achieving impressive scores of 64.5 on LongVideoBench and 35.1 on MMLongBench-Doc. Its native-resolution vision encoder, MoonViT, further allows it to see and understand ultra-high-resolution visual inputs, achieving 83.2 on InfoVQA and 34.5 on ScreenSpot-Pro, while maintaining lower computational cost for common tasks. Building upon Kimi-VL, we introduce an advanced long-thinking variant: Kimi-VL-Thinking-2506. Developed through long chain-of-thought (CoT) supervised fine-tuning (SFT) and reinforcement learning (RL), the latest model exhibits strong long-horizon reasoning capabilities (64.0 on MMMU, 46.3 on MMMU-Pro, 56.9 on MathVision, 80.1 on MathVista, 65.2 on VideoMMMU) while obtaining robust general abilities (84.4 on MMBench, 83.2 on V* and 52.8 on ScreenSpot-Pro). With only around 3B activated parameters, it sets a new standard for efficient yet capable multimodal thinking models. Code and models are publicly accessible at https://github.com/MoonshotAI/Kimi-VL. <details> <summary>x1.png Details</summary>  ### Visual Description ## Scatter Plot: Model Performance vs. Parameter Count ### Overview The image is a scatter plot comparing the performance of various AI models (MathVision Pass@1) against their activated parameter counts (in billions). The plot includes labeled data points, connecting lines for model families, and a legend for color coding. Key trends show a general relationship between parameter count and performance, with notable outliers. --- ### Components/Axes - **X-axis**: "Activated Parameters (B)" with ticks at 3, 10, 30, 70, and 80 billion parameters. - **Y-axis**: "MathVision Pass@1" with ticks from 15 to 65. - **Legend**: Located on the right, mapping colors to model families: - **Blue**: Kimi - **Purple**: Gemma - **Gray**: Owen - **Green**: QVQ - **Red**: DeepSeek and Llama - **Data Points**: Labeled with model names, parameter counts, and versions (e.g., "Kimi-VL-A3B-Thinking-2506"). - **Lines**: Dashed lines connect data points for model families (e.g., Owen-2.5-VL-3B → Owen-2.5-VL-7B → Owen-2.5-VL-32B). --- ### Detailed Analysis #### Data Points and Trends 1. **Kimi Models**: - **Kimi-VL-A3B-Thinking-2506**: 3B parameters, 55 Pass@1 (top-left). - **Kimi-VL-A3B-Thinking**: 3B parameters, 35 Pass@1 (lower-left). - **Trend**: Horizontal line at 3B parameters, showing no improvement in Pass@1 with parameter count. 2. **Gemma Models**: - **Gemma-3-4B-IT**: 3B parameters, 25 Pass@1. - **Gemma-3-12B-IT**: 10B parameters, 35 Pass@1. - **Gemma-3-27B-IT**: 30B parameters, 35 Pass@1. - **Trend**: Dashed purple line slopes upward, indicating improved performance with parameter count up to 30B. 3. **Owen Models**: - **Owen-2.5-VL-3B**: 3B parameters, 20 Pass@1. - **Owen-2.5-VL-7B**: 10B parameters, 25 Pass@1. - **Owen-2.5-VL-32B**: 30B parameters, 35 Pass@1. - **Owen-2.5-VL-72B**: 70B parameters, 35 Pass@1. - **Trend**: Dashed gray line slopes upward, then plateaus at 35 Pass@1 despite increased parameters. 4. **QVQ Models**: - **QVQ-72B-Preview**: 70B parameters, 35 Pass@1. - **QVQ-Max-Preview**: 80B parameters, 48 Pass@1. - **Trend**: Dashed green line slopes upward, showing significant improvement at 80B parameters. 5. **DeepSeek and Llama**: - **DeepSeek-VL2-A4.5B**: 5B parameters, 18 Pass@1. - **Llama-3.2-11B-Inst.**: 11B parameters, 15 Pass@1. - **Trend**: No connecting lines; isolated points with low Pass@1 scores. --- ### Key Observations 1. **Parameter Efficiency**: - Kimi models achieve high Pass@1 (55) with minimal parameters (3B), suggesting superior efficiency. - QVQ models require 80B parameters to reach 48 Pass@1, indicating lower efficiency. 2. **Performance Plateaus**: - Owen and Gemma models plateau at 35 Pass@1 despite increasing parameters beyond 30B, suggesting diminishing returns. 3. **Outliers**: - **QVQ-Max-Preview**: Highest Pass@1 (48) but requires 80B parameters, making it an outlier in efficiency. - **Kimi-VL-A3B-Thinking-2506**: Highest Pass@1 (55) with only 3B parameters, another efficiency outlier. 4. **Low-Performance Models**: - DeepSeek and Llama models have the lowest Pass@1 scores (15–18) despite moderate parameter counts (5–11B). --- ### Interpretation The data suggests that **parameter count alone does not guarantee performance**. Kimi models demonstrate exceptional efficiency, achieving high Pass@1 scores with minimal parameters. In contrast, QVQ models require significantly more parameters for comparable or lower performance, highlighting potential over-parameterization. The Owen and Gemma models show a trend where increasing parameters improves performance up to a point (30B), after which gains plateau. This implies that architectural innovations (e.g., Kimi’s design) may be more critical than raw parameter count. The low performance of DeepSeek and Llama models suggests limitations in their training or optimization strategies. Overall, the plot underscores the importance of balancing parameter efficiency with model architecture for optimal results. </details> Figure 1: Comparison between Kimi-VL-Thinking-2506 and frontier open-source VLMs, including short-thinking VLMs (e.g. Gemma-3 series, Qwen2.5-VL series) and long-thinking VLMs (QVQ-72B/Max-Preview), on MathVision benchmark. Our model achieves strong multimodal reasoning with just 2.8B LLM activated parameters. <details> <summary>x2.png Details</summary>  ### Visual Description ## Bar Chart: Model Performance Across Benchmarks ### Overview The image is a grouped bar chart comparing the performance of seven AI models across nine benchmarks. Models include Kimi-VL-A3B (blue), Qwen2.5-VL-7B (dark gray), DeepSeek-VL2 (light gray), GPT-4o (black), GPT-4o-mini (white), Llama-3-2-11B-Instruct (tan), and Gemma-3-12B-IT (light tan). Benchmarks are categorized into "General," "OCR," "Multi-Image," "Long Video," "Long Doc," and "Agent" tasks. Scores are represented as percentages on the y-axis (0–90 range). --- ### Components/Axes - **X-Axis**: Benchmarks (e.g., MMMU (val), MMBench-EN-v1.1, InfoVQA, BLINK, LongVideoBench, Video-MME (w/o sub), MMLongBench-Doc, ScreenSpot-Pro, OSWorld (Pass@1)). - **Y-Axis**: Performance scores (0–90, labeled "Score"). - **Legend**: Top-left corner, mapping colors to models. - **Bars**: Grouped by benchmark, with individual bars for each model. Numerical values are displayed atop bars. --- ### Detailed Analysis #### General Benchmarks - **MMMU (val)**: - Kimi-VL-A3B: 57.0 - Qwen2.5-VL-7B: 58.6 - DeepSeek-VL2: 51.1 - GPT-4o: 60.0 - GPT-4o-mini: 48.0 - Llama-3-2-11B-Instruct: 59.6 - Gemma-3-12B-IT: 48.0 - **MMBench-EN-v1.1**: - Kimi-VL-A3B: 83.1 - Qwen2.5-VL-7B: 82.6 - DeepSeek-VL2: 79.6 - GPT-4o: 77.1 - GPT-4o-mini: 65.8 - Llama-3-2-11B-Instruct: 74.6 - Gemma-3-12B-IT: 65.8 #### OCR Benchmarks - **InfoVQA**: - Kimi-VL-A3B: 83.2 - Qwen2.5-VL-7B: 82.6 - DeepSeek-VL2: 78.1 - GPT-4o: 57.9 - GPT-4o-mini: 43.8 - Llama-3-2-11B-Instruct: 34.6 - Gemma-3-12B-IT: 43.8 #### Multi-Image Benchmarks - **BLINK**: - Kimi-VL-A3B: 57.3 - Qwen2.5-VL-7B: 56.4 - DeepSeek-VL2: 53.6 - GPT-4o: 39.8 - GPT-4o-mini: 50.3 - Llama-3-2-11B-Instruct: 50.3 - Gemma-3-12B-IT: 50.3 #### Long Video Benchmarks - **LongVideoBench**: - Kimi-VL-A3B: 64.5 - Qwen2.5-VL-7B: 56.0 - DeepSeek-VL2: 58.2 - GPT-4o: 45.5 - GPT-4o-mini: 51.5 - Llama-3-2-11B-Instruct: 46.0 - Gemma-3-12B-IT: 58.2 - **Video-MME (w/o sub)**: - Kimi-VL-A3B: 67.8 - Qwen2.5-VL-7B: 65.1 - DeepSeek-VL2: 64.8 - GPT-4o: 46.0 - GPT-4o-mini: 46.0 - Llama-3-2-11B-Instruct: 58.2 - Gemma-3-12B-IT: 46.0 #### Long Doc Benchmarks - **MMLongBench-Doc**: - Kimi-VL-A3B: 35.1 - Qwen2.5-VL-7B: 29.6 - DeepSeek-VL2: 29.0 - GPT-4o: 13.8 - GPT-4o-mini: 21.3 - Llama-3-2-11B-Instruct: 21.3 - Gemma-3-12B-IT: 21.3 #### Agent Benchmarks - **ScreenSpot-Pro**: - Kimi-VL-A3B: 34.5 - Qwen2.5-VL-7B: 29.0 - DeepSeek-VL2: 0.8 - GPT-4o: 0.8 - GPT-4o-mini: 0.8 - Llama-3-2-11B-Instruct: 0.8 - Gemma-3-12B-IT: 0.8 - **OSWorld (Pass@1)**: - Kimi-VL-A3B: 8.2 - Qwen2.5-VL-7B: 2.5 - DeepSeek-VL2: 5.0 - GPT-4o: 5.0 - GPT-4o-mini: 5.0 - Llama-3-2-11B-Instruct: 5.0 - Gemma-3-12B-IT: 5.0 --- ### Key Observations 1. **Kimi-VL-A3B** consistently outperforms other models in most benchmarks, particularly in **MMBench-EN-v1.1** (83.1) and **Video-MME** (67.8). 2. **GPT-4o** and **GPT-4o-mini** show strong performance in **General** and **OCR** tasks but lag in **Long Doc** and **Agent** benchmarks. 3. **Llama-3-2-11B-Instruct** and **Gemma-3-12B-IT** perform comparably in **Multi-Image** tasks but underperform in **Long Doc** and **Agent** tasks. 4. **DeepSeek-VL2** excels in **LongVideoBench** (58.2) but struggles in **Agent** tasks (0.8). 5. **Qwen2.5-VL-7B** shows balanced performance across most benchmarks but lags in **Agent** tasks. --- ### Interpretation The chart highlights **Kimi-VL-A3B** as the most versatile model, excelling across diverse tasks. **GPT-4o** and **GPT-4o-mini** dominate **General** and **OCR** tasks but underperform in **Long Doc** and **Agent** benchmarks, suggesting limitations in handling extended text or interactive tasks. **Llama-3-2-11B-Instruct** and **Gemma-3-12B-IT** show niche strengths in **Multi-Image** tasks but lack consistency elsewhere. The stark drop in performance for **Agent** tasks (e.g., **ScreenSpot-Pro**: 0.8 for most models) indicates a critical gap in real-world application readiness for many models. This data underscores the importance of model selection based on specific use cases, with **Kimi-VL-A3B** emerging as a strong candidate for broad applicability. </details> Figure 2: Highlights of Kimi-VL performance for a wide range of benchmarks like, general benchmarks (MMMU, MMBench), OCR (InfoVQA), multi-image (BLINK), long video (LongVideoBench, Video-MME), long document (MMLongBench-Doc), and agent (ScreenSpot-Pro and OSWorld). Detailed results are presented in Table 3. 1 Introduction With the rapid advancement of artificial intelligence, human expectations for AI assistants have transcended traditional language-only interactions, increasingly aligning with the inherently multimodal nature of our world. To better understand and interact with these expectations, new generations of natively multimodal models, such as GPT-4o \parencite openai2024gpt4ocard and Google Gemini \parencite geminiteam2024gemini15unlockingmultimodal, have emerged with the capability to seamlessly perceive and interpret visual inputs alongside language processing. Most recently, advanced multimodal models, pioneered by OpenAI o1 series \parencite o12024 and Kimi k1.5 \parencite team2025kimi, have further pushed these boundaries by incorporating deeper and longer reasoning on multimodal inputs, thereby tackling more complex problems in the multimodal domain. Nevertheless, development in large VLMs in the open-source community has significantly lagged behind their language-only counterparts, particularly in aspects of scalability, computational efficiency, and advanced reasoning capabilities. While language-only model DeepSeek R1 \parencite deepseekai2025deepseekr1incentivizingreasoningcapability has already leveraged the efficient and more scalable mixture-of-experts (MoE) architecture and facilitated sophisticated long chain-of-thought (CoT) reasoning, most recent open-source VLMs, e.g. Qwen2.5-VL \parencite bai2025qwen25vltechnicalreport and Gemma-3 \parencite gemmateam2025gemma3technicalreport, continue to rely on dense architectures and do not support long-CoT reasoning. Early explorations into MoE-based vision-language models, such as DeepSeek-VL2 \parencite wu2024deepseekvl2mixtureofexpertsvisionlanguagemodels and Aria \parencite li2024ariaopenmultimodalnative, exhibit limitations in other crucial dimensions. Architecturally, both models still adopt relatively traditional fixed-size vision encoders, hindering their adaptability to diverse visual inputs. From a capability perspective, DeepSeek-VL2 supports only a limited context length (4K), while Aria falls short in fine-grained visual tasks. Additionally, neither of them supports long-thinking abilities. Consequently, there remains a pressing need for an open-source VLM that effectively integrates structural innovation, stable capabilities, and enhanced reasoning through long-thinking. In light of this, we present Kimi-VL, a vision-language model for the open-source community. Structurally, Kimi-VL consists of our Moonlight \parencite liu2025muonscalablellmtraining MoE language model with only 2.8B activated (16B total) parameters, paired with a 400M native-resolution MoonViT vision encoder. In terms of capability, as illustrated in Figure 2, Kimi-VL can robustly handle diverse tasks (fine-grained perception, math, college-level problems, OCR, agent, etc.) across a broad spectrum of input forms (single-image, multi-image, video, long-document, etc.). Specifically, it features the following exciting abilities: 1) Kimi-VL is smart: it has comparable text ability against efficient pure-text LLMs; without long thinking, Kimi-VL is already competitive in multimodal reasoning and multi-turn agent benchmarks, e.g., MMMU, MathVista, OSWorld. 2) Kimi-VL processes long: it effectively tackles long-context understanding on various multimodal inputs within its 128K context window, far ahead of similar-scale competitors on long video benchmarks and MMLongBench-Doc. 3) Kimi-VL perceives clear: it shows all-round competitive ability over existing efficient dense and MoE VLMs in various vision-language scenarios: visual perception, visual world knowledge, OCR, high-resolution OS screenshot, etc. Furthermore, with long-CoT activation and reinforcement learning (RL), we introduce the long-thinking version of Kimi-VL, Kimi-VL-Thinking, which further substantially improves performance on more complex multimodal reasoning scenarios. Despite its small scale, Kimi-VL-Thinking offers compelling performance on hard reasoning benchmarks (e.g., MMMU, MathVision, MathVista), outperforming many state-of-the-art VLMs with even larger sizes. We further release and improved version of the thinking model, Kimi-VL-Thinking-2506. The improved version has even better performance on these reasoning benchmarks while retaining or improving on common visual perception and understanding scenarios, e.g. high-resolution perception (V*), OS grounding, video and long document understanding. 2 Approach 2.1 Model Architecture <details> <summary>x3.png Details</summary>  ### Visual Description ## Diagram: Multimodal Processing Architecture with Mixture-of-Experts and MoonViT ### Overview The diagram illustrates a technical architecture for multimodal processing, combining text and image analysis. It features a Mixture-of-Experts (MoE) Language Decoder, an MLP Projector, and MoonViT (a vision transformer) with native-resolution processing. The system integrates long video, fine-grained images, and UI screenshots, with OCR text extraction as output. ### Components/Axes 1. **Top Section (MoE Language Decoder)**: - **MoE FFN**: Feed-forward neural network with mixture-of-experts. - **Attention Layer**: Standard transformer attention mechanism. - **Router**: Distributes inputs to shared/non-shared experts. - **Experts**: Labeled as "Shared Experts" and "Non-shared Experts" with color-coded connections (red, blue, green, orange). 2. **Middle Section (MLP Projector)**: - Connects MoE Decoder to MoonViT. 3. **Bottom Section (MoonViT)**: - **Input Sources**: - **Long Video**: 270px × 480px (stacked frames). - **Fine-grained Image**: 1008px × 672px (detailed landscape). - **UI Screenshot**: 800px × 1731px (mobile interface). - **Output**: OCR text ("Yas test? That is the exciting competition going on"). 4. **Textual Elements**: - **Input Prompt**: "What can you interpret from..." - **Output Text**: OCR-extracted text from MoonViT. ### Detailed Analysis - **MoE Decoder**: - The Router directs inputs to experts based on query complexity (e.g., red arrows for high-priority tasks). - Non-shared experts (orange) handle specialized tasks, while shared experts (blue/green) manage common operations. - **MoonViT**: - Processes inputs at varying resolutions (20px, 50px, 480px, 1113px, 1008px, 800px). - Outputs OCR text, suggesting text extraction from images (e.g., "CULTURAL CROSSINGS" from the long video). - **UI Screenshot**: - Contains app icons (Calendar, Photos, Camera) and system info (weather: 33°C in Manila). ### Key Observations 1. **Multimodal Integration**: Combines text decoding (MoE) with vision processing (MoonViT). 2. **Resolution Handling**: MoonViT adapts to diverse image sizes (e.g., 20px for small images, 1008px for fine-grained details). 3. **OCR Capability**: Explicitly extracts text from images (e.g., "CULTURAL CROSSINGS" and "Yas test?"). ### Interpretation - **Efficiency**: MoE architecture reduces computational costs by activating only relevant experts. - **Adaptability**: MoonViT’s native-resolution processing handles diverse visual inputs (e.g., UI elements, natural scenes). - **Text-Image Link**: OCR output demonstrates the system’s ability to bridge visual and textual data, useful for tasks like document analysis or UI automation. - **Potential Use Cases**: Multimodal chatbots, video captioning, or UI element recognition. ## Notes - No numerical data or trends present; focus is on architectural components and workflow. - All text is in English; no foreign languages detected. - Spatial grounding confirms connections via color-coded arrows (e.g., red for high-priority routing). </details> Figure 3: The model architecture of Kimi-VL and Kimi-VL-Thinking, consisting of a MoonViT that allows native-resolution images, an MLP projector, and a Mixture-of-Experts (MoE) language decoder. The architecture of Kimi-VL consists of three parts: a native-resolution vision encoder (MoonViT), an MLP projector, and an MoE language model, as depicted in Figure 3. We introduce each part in this section. MoonViT: A Native-resolution Vision Encoder We design MoonViT, the vision encoder of Kimi-VL, to natively process images at their varying resolutions, eliminating the need for complex sub-image splitting and splicing operations, as employed in LLaVA-OneVision \parencite li2024llavaonevisioneasyvisualtask. We incorporate the packing method from NaViT \parencite dehghani2023patchnpacknavit, where images are divided into patches, flattened, and sequentially concatenated into 1D sequences. These preprocessing operations enable MoonViT to share the same core computation operators and optimization as a language model, such as the variable-length sequence attention mechanism supported by FlashAttention \parencite dao2022flashattentionfastmemoryefficientexact, ensuring non-compromised training throughput for images of varying resolutions. MoonViT is initialized from and continually pre-trained on SigLIP-SO-400M \parencite zhai2023sigmoidlosslanguageimage, which originally employs learnable fixed-size absolute positional embeddings to encode spatial information. While we interpolate these original position embeddings to better preserve SigLIP’s capabilities, these interpolated embeddings become increasingly inadequate as image resolution increases. To address this limitation, we incorporate 2D rotary positional embedding (RoPE) \parencite su2023roformerenhancedtransformerrotary across the height and width dimensions, which improves the representation of fine-grained positional information, especially in high-resolution images. These two positional embedding approaches work together to encode spatial information for our model and seamlessly integrate with the flattening and packing procedures. This integration enables MoonViT to efficiently process images of varying resolutions within the same batch. The resulting continuous image features are then forwarded to the MLP projector and, ultimately, to the MoE language model for subsequent training stages. In Kimi-VL-A3B-Thinking-2506, we further continually train this MoonViT to authentically encode up to 3.2 million pixels from a single image, 4 times compared to the original limit. MLP Projector We employ a two-layer MLP to bridge the vision encoder (MoonViT) and the LLM. Specifically, we first use a pixel shuffle operation to compress the spatial dimension of the image features extracted by MoonViT, performing 2×2 downsampling in the spatial domain and correspondingly expanding the channel dimension. We then feed the pixel-shuffled features into a two-layer MLP to project them into the dimension of LLM embeddings. Mixture-of-Experts (MoE) Language Model The language model of Kimi-VL utilizes our Moonlight model \parencite liu2025muonscalablellmtraining, an MoE language model with 2.8B activated parameters, 16B total parameters, and an architecture similar to DeepSeek-V3 \parencite deepseekai2025deepseekv3technicalreport. For our implementation, we initialize from an intermediate checkpoint in Moonlight’s pre-training stage—one that has processed 5.2T tokens of pure text data and activated an 8192-token (8K) context length. We then continue pre-training it using a joint recipe of multimodal and text-only data totaling 2.3T tokens, as detailed in Sec. 2.3. 2.2 Muon Optimizer We use an enhanced Muon optimizer \parencite liu2025muon for model optimization. Compared to the original Muon optimizer \parencite jordan2024muon, we add weight decay and carefully adjust the per-parameter update scale. Additionally, we develop a distributed implementation of Muon following the ZeRO-1 \parencite rajbhandari2020zero optimization strategy, which achieves optimal memory efficiency and reduced communication overhead while preserving the algorithm’s mathematical properties. This enhanced Muon optimizer is used throughout the entire training process to optimize all model parameters, including the vision encoder, the projector, and the language model. 2.3 Pre-Training Stages As illustrated in Figure 4 and Table 1, after loading the intermediate language model discussed above, Kimi-VL’s pre-training comprises a total of 4 stages consuming 4.4T tokens overall: first, standalone ViT training to establish a robust native-resolution visual encoder, followed by three joint training stages (pre-training, cooldown, and long-context activation) that simultaneously enhance the model’s language and multimodal capabilities. The details are as follows. <details> <summary>x4.png Details</summary>  ### Visual Description ## Flowchart: Multimodal Model Training Pipeline ### Overview The image depicts a four-stage training pipeline for a multimodal AI model, visualized as a horizontal flowchart with blue rectangles connected by green arrows. Each stage represents a distinct training phase with specific data requirements and objectives. The flow progresses from left to right, with green arrows labeled "resumes LR scheduler" indicating iterative optimization steps. ### Components/Axes 1. **Stages (Left to Right):** - Text Pre-training - ViT Training - Joint Pre-training - Joint Cooldown - Joint Long-context 2. **Data Sizes:** Expressed in terabytes (T), with ranges (e.g., 2.0T -> 0.1T) indicating progressive reduction. 3. **Key Attributes:** - Data type (text, multimodal, long-context) - Technical specifications (e.g., CoCa-loss, RoPE base) - Optimization strategies (e.g., LR scheduler resumption) ### Detailed Analysis 1. **Text Pre-training** - **Data:** 5.2T pure text - **Purpose:** Foundational language understanding - **Color:** Light blue 2. **ViT Training** - **Data:** 2.0T -> 0.1T (text-to-image alignment) - **Technique:** CoCa-loss with tiny language decoder - **Objective:** Align vision-language representations - **Color:** Dark blue 3. **Joint Pre-training** - **Data:** 1.4T (40% multimodal) - **Feature:** Progressive multimodal ratio - **Color:** Medium blue 4. **Joint Cooldown** - **Data:** 0.6T (high-quality text/multimodal) - **Strategy:** Re-warmup to higher learning rate (LR) - **Color:** Dark blue 5. **Joint Long-context** - **Data:** 0.3T (long text/video/docs) - **Technical:** RoPE base increased from 50k -> 800k - **Color:** Light blue ### Key Observations - **Data Progression:** Total data decreases from 5.2T to 0.3T, but complexity increases (text → multimodal → long-context). - **Optimization Cycles:** LR scheduler resumption between stages suggests iterative refinement. - **Multimodal Focus:** 40% multimodal data in Joint Pre-training highlights hybrid training emphasis. - **Context Scaling:** RoPE base expansion (50k → 800k) indicates specialized long-sequence handling. ### Interpretation This pipeline demonstrates a staged approach to multimodal model development: 1. **Foundation Building:** Start with massive text pre-training (5.2T) to establish language understanding. 2. **Vision Integration:** Use ViT training (2.0T→0.1T) with CoCa-loss to bridge text-image gaps. 3. **Hybrid Optimization:** Joint Pre-training (1.4T) introduces multimodal data with progressive ratio adjustments. 4. **Efficiency Phase:** Cooldown (0.6T) focuses on high-quality data while re-warming LR for stability. 5. **Specialization:** Final stage (0.3T) targets long-context understanding via RoPE base expansion. The decreasing data sizes with increasing technical complexity suggest a shift from broad data collection to targeted, high-quality training. The LR scheduler resumption between stages implies careful learning rate management to avoid overfitting while maintaining model adaptability. The 40% multimodal ratio in Joint Pre-training indicates a balanced approach to hybrid training before final specialization in long-context scenarios. </details> Figure 4: The pre-training stages of Kimi-VL consume a total of 4.4T tokens after text-only pre-training of its language model. To preserve text abilities, all stages that update the language model are joint training stages. Table 1: Overview of training stages: data composition, token volumes, sequence lengths, and trainable components. | Stages Data | ViT Training Alt text Synthesis Caption | Joint Pre-training + Text, Knowledge | Joint Cooldown + High-quality Text | Joint Long-context + Long Text | | --- | --- | --- | --- | --- | | Grounding | Interleaving | High-quality Multimodal | Long Video | | | OCR | Video, Agent | Academic Sources | Long Document | | | Tokens | 2T + 0.1T | 1.4T | 0.6T | 0.3T | | Sequence length | 8192 | 8192 | 8192 | 32768->131072 | | Training | ViT | ViT & LLM | ViT & LLM | ViT & LLM | ViT Training Stages The MoonViT is trained on image-text pairs, where the text components consist of a variety of targets: image alt texts, synthetic captions, grounding bboxes, and OCR texts. The training incorporates two objectives: a SigLIP \parencite zhai2023sigmoidlosslanguageimage loss $\mathcal{L}_{siglip}$ (a variant of contrastive loss) and a cross-entropy loss $\mathcal{L}_{caption}$ for caption generation conditioned on input images. Following CoCa’s approach \parencite yu2022cocacontrastivecaptionersimagetext, the final loss function is formulated as $\mathcal{L}=\mathcal{L}_{siglip}+\lambda\mathcal{L}_{caption}$ , where $\lambda=2$ . Specifically, the image and text encoders compute the contrastive loss, while the text decoder performs next-token prediction (NTP) conditioned on features from the image encoder. To accelerate training, we initialized both encoders with SigLIP SO-400M \parencite zhai2023sigmoidlosslanguageimage weights and implemented a progressive resolution sampling strategy to gradually allow larger size; the text decoder is initialized from a tiny decoder-only language model. During training, we observed an emergence in the caption loss while scaling up OCR data, indicating that the text decoder had developed some OCR capabilities. After training the ViT in the CoCa-alike stage with 2T tokens, we align the MoonViT to the MoE language model using another 0.1T tokens, where only MoonViT and MLP projector are updated. This alignment stage significantly reduces the initial perplexity of MoonViT embeddings in the language model, allowing a smoother joint pre-training stage as follows. Joint Pre-training Stage In the joint pre-training stage, we train the model with a combination of pure text data (sampled from the same distribution as the initial language model) and a variety of multimodal data (as discussed in Sec. 3.1). We continue training from the loaded LLM checkpoint using the same learning rate scheduler, consuming an additional 1.4T tokens. The initial steps utilize solely language data, after which the proportion of multimodal data gradually increases. Through this progressive approach and the previous alignment stage, we observe that joint pre-training preserves the model’s language capabilities while successfully integrating visual comprehension abilities. Joint Cooldown Stage The stage following the pre-training stage is a multimodal cooldown phase, where the model is continue trained with high-quality language and multimodal datasets to ensure superior performance. For the language part, through empirical investigation, we observe that the incorporation of synthetic data during the cooling phase yields significant performance improvements, particularly in mathematical reasoning, knowledge-based tasks, and code generation. The general text components of the cooldown dataset are curated from high-fidelity subsets of the pre-training corpus. For math, knowledge, and code domains, we employ a hybrid approach: utilizing selected pre-training subsets while augmenting them with synthetically generated content. Specifically, we leverage existing mathematical knowledge and code corpora as source material to generate question-answer (QA) pairs through a proprietary language model, implementing rejection sampling techniques to maintain quality standards \parencite yue2023mammoth,su2024nemotron. These synthesized QA pairs undergo comprehensive validation before being integrated into the cooldown dataset. For the multimodal part, in addition to the two strategies as employed in text cooldown data preparation, i.e. question-answer synthesis and high-quality subset replay, to allow more comprehensive visual-centric perception and understanding \parencite li2024llavaonevisioneasyvisualtask,tong2024cambrian1fullyopenvisioncentric,guo2024mammothvlelicitingmultimodalreasoning, we filter and rewrite a variety of academic visual or vision-language data sources to QA pairs. Unlike post-training stages, these language and multimodal QA pairs in the cooldown stage are only included for activating specific abilities and henceforth facilitating learning high-quality data, thus, we keep their ratio at a low portion to avoid overfitting these QA patterns. The joint cooldown stage significantly improves both language and multimodal abilities of the model. Table 2: Needle-in-a-Haystack (NIAH) test on text/video haystacks, where needles are uniformly distributed at various positions within the haystack. We report recall accuracy across different haystack lengths up to 131,072 tokens (128K). | - text haystack - video haystack | 100.0 100.0 | 100.0 100.0 | 100.0 100.0 | 100.0 100.0 | 100.0 100.0 | 100.0 100.0 | 87.0 91.7 | | --- | --- | --- | --- | --- | --- | --- | --- | Joint Long-context Activation Stage In the final pre-training stage, we extend the context length of the model from 8192 (8K) to 131072 (128K), with the inverse frequency of its RoPE \parencite su2023roformerenhancedtransformerrotary embeddings reset from 50,000 to 800,000. The joint long-context stage is conducted in two sub-stages, where each one extends the model’s context length by four times. For data composition, we filter and upsample the ratio of long data to 25% in each sub-stage, while using the remaining 75% tokens to replay shorter data in its previous stage; our exploration confirms that this composition allows the model to effectively learn long-context understanding while maintaining short-context ability. To allow the model to activate long-context abilities on both pure-text and multimodal inputs, the long data used in Kimi-VL’s long-context activation consists of not only long text, but also long multimodal data, including long interleaved data, long videos, and long documents. Similar as cooldown data, we also synthesize a small portion of QA pairs to augment the learning efficiency of long-context activation. After the long-context activations, the model can pass needle-in-a-haystack (NIAH) evaluations with either long pure-text or long video haystack, proving its versatile long-context ability. We provide the NIAH recall accuracy on various range of context length up to 128K in Table 2. <details> <summary>x5.png Details</summary>  ### Visual Description ## Flowchart: Kimi-VL Training Pipeline ### Overview The image depicts a three-stage training pipeline for the Kimi-VL model, visualized as a horizontal flowchart with three blue rectangular blocks connected by arrows. Each block represents a distinct training phase, with technical specifications and objectives labeled in white text. Arrows between blocks indicate progression and data flow. ### Components/Axes 1. **Blocks (Training Phases)**: - **Left Block**: "Joint Supervised Fine-tuning" Subtext: "Text + Multimodal SFT Data\n1 Epoch@32K + 1 Epoch@128K" - **Middle Block**: "Long-CoT Supervised Fine-tuning" Subtext: "Text + Multimodal Long-CoT Data\nPlanning, Evaluation, Reflection, Exploration" - **Right Block**: "Reinforcement Learning (RL)" Subtext: "Online RL on Answer Only\nLength penalty, Difficulty control" 2. **Arrows (Flow Direction)**: - Left-to-middle arrow labeled "Kimi-VL" - Middle-to-right arrow labeled "Kimi-VL-Thinking" ### Detailed Analysis - **Block 1 (Joint Supervised Fine-tuning)**: Focuses on initial training using text and multimodal data with two distinct training epochs (32K and 128K tokens). The term "SFT" (Supervised Fine-Tuning) implies standard supervised learning with labeled data. - **Block 2 (Long-CoT Supervised Fine-tuning)**: Builds on the first phase by incorporating Chain-of-Thought (CoT) data, emphasizing reasoning capabilities through explicit subtext categories: planning, evaluation, reflection, and exploration. This suggests a focus on developing structured reasoning processes. - **Block 3 (Reinforcement Learning)**: Shifts to online RL with answer-only feedback, introducing constraints like length penalties and difficulty control. This phase prioritizes optimizing response quality and efficiency. ### Key Observations 1. **Progressive Complexity**: Each phase adds specialized components (CoT data, RL constraints) to enhance the model's capabilities. 2. **Data Flow**: The pipeline transitions from supervised learning (Blocks 1-2) to reinforcement learning (Block 3), indicating a hybrid approach. 3. **Technical Specificity**: Epoch sizes (32K, 128K) and constraints (length penalty) are explicitly quantified, suggesting rigorous experimentation. ### Interpretation The flowchart illustrates a staged methodology for training Kimi-VL, where each phase addresses specific limitations of the prior. The progression from supervised fine-tuning to RL reflects a common pattern in LLM development: - **Phase 1** establishes foundational knowledge via SFT. - **Phase 2** enhances reasoning via CoT data, critical for complex tasks. - **Phase 3** optimizes real-world performance through RL, balancing answer quality and efficiency. The explicit mention of "Answer Only" in RL suggests a focus on response generation rather than multimodal outputs, while "Difficulty control" implies adaptive training for varying task complexities. This pipeline likely aims to balance breadth (multimodal data) and depth (reasoning/RL) in Kimi-VL's capabilities. </details> Figure 5: The post-training stages of Kimi-VL and Kimi-VL-Thinking, including two stages of joint SFT in 32K and 128K context, and further long-CoT SFT and RL stages to activate and enhance long thinking abilities. 2.4 Post-Training Stages Joint Supervised Fine-tuning (SFT) In this phase, we fine-tune the base model of Kimi-VL with instruction-based fine-tuning to enhance its ability to follow instructions and engage in dialogue, culminating in the creation of the interactive Kimi-VL model. This is achieved by employing the ChatML format (Openai, 2024), which allows for a targeted instruction optimization while maintaining architectural consistency with Kimi-VL. We optimize the language model, MLP projector, and vision encoder using a mixture of pure-text and vision-language SFT data, which will be described in Sec 3.2. Supervision is applied only to answers and special tokens, with system and user prompts being masked. The model is exposed to a curated set of multimodal instruction-response pairs, where explicit dialogue role tagging, structured injection of visual embeddings, and preservation of cross-modal positional relationships are ensured through the format-aware packing. Additionally, to guarantee the model’s comprehensive proficiency in dialogue, we incorporate a mix of multimodal data and pure text dialogue data used in Moonlight, ensuring its versatility across various dialogue scenarios. We first train the model at the sequence length of 32k tokens for 1 epoch, followed by another epoch at the sequence length of 128k tokens. In the first stage (32K), the learning rate decays from $2× 10^{-5}$ to $2× 10^{-6}$ , before it re-warmups to $1× 10^{-5}$ in the second stage (128K) and finally decays to $1× 10^{-6}$ . To improve training efficiency, we pack multiple training examples into each single training sequence. Long-CoT Supervised Fine-Tuning With the refined RL prompt set, we employ prompt engineering to construct a small yet high-quality long-CoT warmup dataset, containing accurately verified reasoning paths for both text and image inputs. This approach resembles rejection sampling (RS) but focuses on generating long-CoT reasoning paths through prompt engineering. The resulting warmup dataset is designed to encapsulate key cognitive processes that are fundamental to human-like reasoning, such as planning, where the model systematically outlines steps before execution; evaluation, involving critical assessment of intermediate steps; reflection, enabling the model to reconsider and refine its approach; and exploration, encouraging consideration of alternative solutions. By performing a lightweight SFT on this warm-up dataset, we effectively prime the model to internalize these multimodal reasoning strategies. As a result, the fine-tuned long-CoT model demonstrates improved capability in generating more detailed and logically coherent responses, which enhances its performance across diverse reasoning tasks. Reinforcement Learning To further advance the model’s reasoning abilities, we then train the model with reinforcement learning (RL), enabling the model to autonomously generate structured CoT rationales. Specifically, similar as Kimi k1.5 \parencite team2025kimi, we adopt a variant of online policy mirror descent as our RL algorithm, which iteratively refines the policy model $\pi_{\theta}$ to improve its problem-solving accuracy. During the $i$ -th training iteration, we treat the current model as a reference policy model and optimize the following objective, regularized by relative entropy to stabilize policy updates: $$ \displaystyle\max_{\theta}\mathbb{E}_{(x,y^{*})\sim\mathcal{D}}\left[\mathbb{E% }_{(y,z)\sim\pi_{\theta}}\left[r(x,y,y^{*})\right]-\tau\mathrm{KL}(\pi_{\theta% }(x)||\pi_{\theta_{i}}(x))\right]\,, \tag{1} $$ where $r$ is a reward model that justifies the correctness of the proposed answer $y$ for the given problem $x$ , by assigning a value $r(x,y,y^{*})∈\{0,1\}$ based on the ground truth $y^{*}$ , and $\tau>0$ is a parameter controlling the degree of regularization. Each training iteration begins by sampling a problem batch from the dataset $\mathcal{D}$ , and the model parameters are updated to $\theta_{i+1}$ using the policy gradient derived from (1), with the optimized policy model subsequently assuming the role of reference policy for the subsequent iteration. To enhance RL training efficiency, we implement a length-based reward to penalize excessively long responses, mitigating the overthinking problem where the model generates redundant reasoning chains. Besides, we employ two sampling strategies including curriculum sampling and prioritized sampling, which leverage difficulty labels and per-instance success rates to focus training effort on the most pedagogically valuable examples, thereby optimizing the learning trajectory and improving training efficiency. Through large-scale reinforcement learning training, we can derive a model that harnesses the strengths of both basic prompt-based CoT reasoning and sophisticated planning-enhanced CoT approaches. During inference, the model maintains standard autoregressive sequence generation, eliminating the deployment complexities associated with specialized planning algorithms that require parallel computation. Simultaneously, the model develops essential meta-reasoning abilities including error detection, backtracking, and iterative solution refinement by effectively utilizing the complete history of explored reasoning paths as contextual information. With endogenous learning from its complete reasoning trace history, the model can effectively encode planned search procedures into its parametric knowledge. 2.5 Infrastructure Storage We utilize S3 \parencite amazon_s3 compatible object storage from cloud service vendors to store our visual-text data. To minimize the time between data preparation and model training, we store visual data in its original format and have developed an efficient and flexible data loading system. This system provides several key benefits: - Supports on-the-fly data shuffling, mixing, tokenization, loss masking and packing during training, allowing us to adjust data proportions as needed; - Enables random augmentation of both visual and text data, while preserving the correctness of 2D coordinate and orientation information during transformations; - Ensures reproducibility by strictly controlling random states and other states across different data loader workers, guaranteeing that any interrupted training can be resumed seamlessly—the data sequence after resumption remains identical to an uninterrupted run; - Delivers high-performance data loading: through multiple caching strategies, our system reliably supports training on large scale clusters while maintaining controlled request rates and throughput to the object storage. Additionally, to ensure consistent dataset quality control, we developed a centralized platform for data registration, visualization, compiling statistics, synchronizing data across cloud storage systems, and managing dataset lifecycles. <details> <summary>x6.png Details</summary>  ### Visual Description ## Screenshot: Manuscript Analysis Response ### Overview The image contains a technical response analyzing two handwritten manuscript pages. The analysis focuses on identifying the author and content based on visual and textual clues, concluding the manuscripts belong to Albert Einstein. ### Components/Axes No charts, diagrams, or data tables are present. The image consists of: 1. **Instruction Section**: Text prompting analysis of manuscript ownership. 2. **Response Section**: Structured analysis with key observations, content details, and conclusion. ### Detailed Analysis #### Key Observations - **Handwriting Style**: Consistent with Einstein’s known style (careful, slightly messy, mathematical notation). - **Content Analysis**: - Equations involve partial derivatives, summations, and Greek letters (e.g., γ, g). - References to constants like "K" (gravitational constant), "M" (mass), and "T" (time). - Terms like "Einheitsvektor" (unit vector) and "Gln" (equation) suggest German-language context. - **German Terms**: Use of "Einheitsvektor" and "Gln" indicates German-speaking author. #### Content Details 1. **First Image (Image 1)**: - Equations involve partial derivatives and summations. - Terms like "Einheitsvektor" and references to gravitational fields. - Style matches early 20th-century mathematical notation. 2. **Second Image (Image 2)**: - Numerical calculations, integrals, and algebraic manipulations. - Constants "K," "M," and "T" align with general relativity equations. - Messy but consistent handwriting suggests a single author. 3. **Conclusion**: - Manuscripts document Einstein’s work on gravitational field equations and general relativity. - Likely part of his research materials for formalizing the mathematical framework of general relativity. ### Interpretation The analysis infers the manuscripts belong to Albert Einstein based on: - **Stylistic Consistency**: Handwriting and notation match Einstein’s documented style. - **Content Relevance**: Equations and terms align with his work on general relativity (e.g., Einstein field equations). - **Linguistic Clues**: German terms ("Einheitsvektor") reflect Einstein’s native language. The response emphasizes that the manuscripts are research documents, not final publications, capturing Einstein’s iterative process in developing gravitational field theory. ### Key Observations (Reiterated) - No data visualizations present; analysis relies on textual and stylistic evidence. - German terms and mathematical content are critical identifiers. - The conclusion is probabilistic ("likely belong to Einstein") due to circumstantial evidence. </details> Figure 6: Manuscript reasoning visualization. Kimi-VL-Thinking demonstrates the ability to perform historical and scientific inference by analyzing handwritten manuscripts step by step. In this example, our model identifies the author as Albert Einstein based on handwriting style, content analysis, and language cues. It reasons that the manuscripts relate to gravitational field equations, consistent with Einstein’s contributions to general relativity. Parallelism We adopt a 4D parallelism strategy—Data Parallelism \parencite li2020pytorchdistributedexperiencesaccelerating, Expert Parallelism \parencite fedus2022switchtransformersscalingtrillion, Pipeline Parallelism \parencite huang2019gpipeefficienttraininggiant,narayanan2021efficientlargescalelanguagemodel, and Context Parallelism \parencite jacobs2023deepspeedulyssesoptimizationsenabling,liu2023ringattentionblockwisetransformers—to accelerate the speed of Kimi-VL . After optimizing parallel strategies, the resulting training throughput of our model is around 60% higher than a 7B dense VLM (e.g. VLMs based on Qwen2.5-7B). - Data Parallelism (DP). DP replicates the model across multiple devices, each processing different micro-batches. This setup allows larger effective batch sizes by simply increasing the number of devices. - Expert Parallelism (EP). EP distributes expert modules in the MoE layer across multiple devices. When combined with DP, experts on a given device can handle tokens from different DP groups, enhancing computational efficiency. - Pipeline Parallelism (PP). PP splits the model into multiple layer-based stages. To minimize pipeline bubbles, we allocate the Vision Tower (VT) and several decoder layers to the first stage, place the output layer and additional decoder layers in the last stage, and distribute the remaining decoder layers evenly across intermediate stages based on their time overhead. - Context Parallelism (CP). CP addresses long-sequence training by splitting sequences across different CP ranks in conjunction with flash attention \parencite dao2022flashattentionfastmemoryefficientexact. This substantially reduces peak memory usage and relieves the memory pressure from attention computations. Beyond these four parallel strategies, we incorporate ZeRO1 \parencite rajbhandari2020zero and Selective Checkpointing Activation \parencite chen2016trainingdeepnetssublinear, korthikanti2022reducingactivationrecomputationlarge to further optimize memory usage. ZeRO1 reduces optimizer state overhead by using a distributed optimizer while avoiding extra communication costs. Selective Checkpointing Activation trades time for space by recomputing only those layers that have low time overhead but high memory consumption, striking a balance between computation efficiency and memory demands. For extremely long sequences, we expand recomputation to a broader set of layers to prevent out-of-memory errors. 3 Data Construction 3.1 Pre-Training Data Our multimodal pre-training corpus is designed to provide high-quality data that enables models to process and understand information from multiple modalities, including text, images, and videos. To this end, we have also curated high-quality data from six categories – caption, interleaving, OCR, knowledge, video, and agent – to form the corpus. When constructing our training corpus, we developed several multimodal data processing pipelines to ensure data quality, encompassing filtering, synthesis, and deduplication. Establishing an effective multimodal data strategy is crucial during the joint training of vision and language, as it both preserves the capabilities of the language model and facilitates alignment of knowledge across diverse modalities. We provide a detailed description of these sources in this section, which is organized into the following categories: Caption Data Our caption data provides the model with fundamental modality alignment and a broad range of world knowledge. By incorporating caption data, the multimodal LLM gains wider world knowledge with high learning efficiency. We have integrated various open-source Chinese and English caption datasets like \parencite schuhmann2022laion, gadre2024datacomp and also collected substantial in-house caption data from multiple sources. However, throughout the training process, we strictly limit the proportion of synthetic caption data to mitigate the risk of hallucination stemming from insufficient real-world knowledge. For general caption data, we follow a rigorous quality control pipeline that avoids duplication and maintain high image-text correlation. We also vary image resolution during pre-training to ensure that the vision tower remains effective when processing images of both high- and low-resolution. Image-text Interleaving Data During the pre-training phase, the model benefits from interleaving data for many aspects. For example, multi-image comprehension ability can be boosted by interleaving data; interleaving data always provides detailed knowledge for the given image; a longer multimodal context learning ability can also be gained by interleaving data. What’s more, we also find that interleaving data can contribute positively to maintaining the model’s language abilities. Thus, image-text interleaving data is an important part in our training corpus. Our multimodal corpus considered open-sourced interleave datasets like \parencite zhu2024multimodal,laurenccon2024obelics and also constructed large-scale in-house data using resources like textbooks, webpages, and tutorials. Further, we also find that synthesizing the interleaving data benefits the performance of multimodal LLM for keeping the text knowledge. To ensure each image’s knowledge is sufficiently studied, for all the interleaving data, despite standard filtering, deduping, and other quality control pipeline, we also integrate a data reordering procedure to keep all the image and text in the correct order. OCR Data Optical Character Recognition (OCR) is a widely adopted technique that converts text from images into an editable format. In our model, a robust OCR capability is deemed essential for better aligning the model with human values. Accordingly, our OCR data sources are diverse, ranging from open-source to in-house datasets, encompassing both clean and augmented images, and spanning over single-page and multi-page inputs. In addition to the publicly available data, we have developed a substantial volume of in-house OCR datasets, covering multilingual text, dense text layouts, web-based content, and handwritten samples. Furthermore, following the principles outlined in OCR 2.0 \parencite wei2024general, our model is also equipped to handle a variety of optical image types, including figures, tables, geometry diagrams, mermaid plots, and natural scene text. We apply extensive data augmentation techniques—such as rotation, distortion, color adjustments, and noise addition—to enhance the model’s robustness. As a result, our model achieves a high level of proficiency in OCR tasks. In addition to single-page OCR data, we collect and convert a large volume of in-house multi-page OCR data to activate the model’s understanding of long documents in the real world. With the help of these data, our model is capable of performing accurate OCR on a single image but can also comprehend an entire academic paper or a scanned book. Knowledge Data The concept of multimodal knowledge data is analogous to the previously mentioned text pre-training data, except here we focus on assembling a comprehensive repository of human knowledge from diverse sources to further enhance the model’s capabilities. For example, carefully curated geometry data in our dataset is vital for developing visual reasoning skills, ensuring the model can interpret the abstract diagrams created by humans. Our knowledge corpus adheres to a standardized taxonomy to balance content across various categories, ensuring diversity in data sources. Similar to text-only corpora, which gather knowledge from textbooks, research papers, and other academic materials, multimodal knowledge data employs both a layout parser and an OCR model to process content from these sources. While we also include filtered data from internet-based and other external resources. Because a significant portion of our knowledge corpus is sourced from internet-based materials, infographics can cause the model to focus solely on OCR-based information. In such cases, relying exclusively on a basic OCR pipeline may limit training effectiveness. To address this, we have developed an additional pipeline that better captures the purely textual information embedded within images. Agent Data For agent tasks, the model’s grounding and planning capabilities have been significantly enhanced. In addition to utilizing publicly available data, a platform has been established to efficiently manage and execute virtual machine environments in bulk. Within these virtual environments, heuristic methods were employed to collect screenshots and corresponding action data. This data was then processed into dense grounding formats and continuous trajectory formats. The design of the Action Space was categorized according to Desktop, Mobile, and Web environments. Furthermore, icon data was collected to strengthen the model’s understanding of the meanings of icons within software graphical user interfaces (GUIs). To enhance the model’s planning ability for solving multi-step desktop tasks, a set of computer-use trajectories was collected from human annotators, each accompanied by synthesized Chain-of-Thought (Aguvis \parencite xu2024aguvis). These multi-step agent demonstrations equip Kimi-VL with the capability to complete real-world desktop tasks (on both Ubuntu and Windows). Video Data In addition to image-only and image-text interleaved data, we also incorporate large-scale video data during pre-training, cooldown, and long-context activation stages to enable two directions of essential abilities of our model: first, to understand a long-context sequence dominated by images (e.g. hour-long videos) in addition to long text; second, to perceive fine-grained spatio-temporal correspondence in short video clips. Our video data are sourced from diverse resources, including open-source datasets as well as in-house web-scale video data, and span videos of varying durations. Similarly, to ensure sufficient generalization ability, our video data cover a wide range of scenes and diverse tasks. We cover tasks such as video description and video grounding, among others. For long videos, we carefully design a pipeline to produce dense captions. Similar to processing the caption data, we strictly limit the proportion of the synthetic dense video description data to reduce the risk of hallucinations. Text Data Our text pretrain corpus directly utilizes the data in Moonlight [liu2025muonscalablellmtraining], which is designed to provide comprehensive and high-quality data for training large language models (LLMs). It encompasses five domains: English, Chinese, Code, Mathematics & Reasoning, and Knowledge. We employ sophisticated filtering and quality control mechanisms for each domain to ensure the highest quality training data. For all pretrain data, we conducted rigorous individual validation for each data source to assess its specific contribution to the overall training recipe. This systematic evaluation ensures the quality and effectiveness of our diverse data composition. To optimize the overall composition of our training corpus, the sampling strategy for different document types is empirically determined through extensive experimentation. We conduct isolated evaluations to identify document subsets that contribute most significantly to the model’s knowledge acquisition capabilities. These high-value subsets are upsampled in the final training corpus. However, to maintain data diversity and ensure model generalization, we carefully preserve a balanced representation of other document types at appropriate ratios. This data-driven approach helps us optimize the trade-off between focused knowledge acquisition and broad generalization capabilities. footnotetext: GPT-4o and GPT-4o-mini results use Omniparser without UIA, according to [bonatti2024windowsagentarenaevaluating]. | | Benchmark (Metric) | GPT-4o | GPT- | Qwen2.5- | Llama3.2- | Gemma3- | DeepSeek- | Kimi-VL- | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | 4o-mini | VL-7B | 11B-Inst. | 12B-IT | VL2 | A3B | | | | | Architecture | - | - | Dense | Dense | Dense | MoE | MoE | | | # Act. Params ${}_{\text{(LLM+VT)}}$ | - | - | 7.6B+0.7B | 8B+2.6B | 12B+0.4B | 4.1B+0.4B | 2.8B+0.4B | | | # Total Params | - | - | 8B | 11B | 12B | 28B | 16B | | | College-level | MMMU ${}_{\text{val}}$ (Pass@1) | 69.1 | 60.0 | 58.6 | 48 | 59.6 | 51.1 | 57.0 | | VideoMMMU (Pass@1) | 61.2 | - | 47.4 | 41.8 | 57.2 | 44.4 | 52.6 | | | MMVU ${}_{\text{val}}$ (Pass@1) | 67.4 | 61.6 | 50.1 | 44.4 | 57.0 | 52.1 | 52.2 | | | General | MMBench-EN-v1.1 (Acc) | 83.1 | 77.1 | 82.6 | 65.8 | 74.6 | 79.6 | 83.1 | | MMStar (Acc) | 64.7 | 54.8 | 63.9 | 49.8 | 56.1 | 55.5 | 61.3 | | | MMVet (Pass@1) | 69.1 | 66.9 | 67.1 | 57.6 | 64.9 | 60.0 | 66.7 | | | RealWorldQA (Acc) | 75.4 | 67.1 | 68.5 | 63.3 | 59.1 | 68.4 | 68.1 | | | AI2D (Acc) | 84.6 | 77.8 | 83.9 | 77.3 | 78.1 | 81.4 | 84.9 | | | Multi-image | BLINK (Acc) | 68.0 | 53.6 | 56.4 | 39.8 | 50.3 | - | 57.3 | | Math | MathVista (Pass@1) | 63.8 | 52.5 | 68.2 | 47.7 | 56.1 | 62.8 | 68.7 | | MathVision (Pass@1) | 30.4 | - | 25.1 | 13.6 | 32.1 | 17.3 | 21.4 | | | OCR | InfoVQA (Acc) | 80.7 | 57.9 | 82.6 | 34.6 | 43.8 | 78.1 | 83.2 | | OCRBench (Acc) | 815 | 785 | 864 | 753 | 702 | 811 | 867 | | | OS Agent | ScreenSpot-V2 (Acc) | 18.1 | - | 86.8 | - | - | - | 92.8 | | ScreenSpot-Pro (Acc) | 0.8 | - | 29.0 | - | - | - | 34.5 | | | OSWorld (Pass@1) | 5.03 | - | 2.5 | - | - | - | 8.22 | | | WindowsAgentArena (Pass@1) footnotemark: | 9.4 | 2.7 | 3.4 | - | - | - | 10.4 | | | Long Document | MMLongBench-Doc (Acc) | 42.8 | 29.0 | 29.6 | 13.8 | 21.3 | - | 35.1 | | Long Video | Video-MME (w/o sub. / w/ sub.) | 71.9/77.2 | 64.8/68.9 | 65.1/71.6 | 46.0/49.5 | 58.2/62.1 | - | 67.8/72.6 | | MLVU ${}_{\text{MCQ}}$ (Acc) | 64.6 | 48.1 | 70.2 | 44.4 | 52.3 | - | 74.2 | | | LongVideoBench ${}_{\text{val}}$ | 66.7 | 58.2 | 56.0 | 45.5 | 51.5 | - | 64.5 | | | Video Perception | EgoSchema ${}_{\text{full}}$ | 72.2 | - | 65.0 | 54.3 | 56.9 | 38.5 | 78.5 | | VSI-Bench | 34.0 | - | 34.2 | 20.6 | 32.4 | 21.7 | 37.4 | | | TOMATO | 37.7 | 28.8 | 27.6 | 21.5 | 28.6 | 27.2 | 31.7 | | Table 3: Performance of Kimi-VL against proprietary and open-source efficient VLMs; performance of GPT-4o is also listed in gray for reference. Top and second-best models are in boldface and underline respectively. Some results of competing models are unavailable due to limitation of model ability on specific tasks or model context length. 3.2 Instruction Data At this stage, the data is primarily aimed at enhancing the model’s conversational abilities and instruction-following capabilities. To cover as many scenarios as possible, we enrich the data across different domains. For non-reasoning tasks, including chart interpretation, agent grounding, OCR, image-grounded conversations, question-answering, writing, and text processing, we initially construct a seed dataset through human annotation. This seed dataset is used to train a seed model. Subsequently, we collect a diverse set of prompts and employ the seed model to generate multiple responses to each prompt. Annotators then rank these responses and refine the top-ranked response to produce the final version. For reasoning tasks like visual coding, visual reasoning, and math/science problems, where rule-based and model-based verifications are more accurate and efficient than human judgment, we utilize rejection sampling to expand the SFT dataset. The complete vanilla SFT dataset comprises approximately a 1:1 ratio of text tokens to image tokens. <details> <summary>x7.png Details</summary>  ### Visual Description ## Screenshot: Multi-Panel Q&A Interface with Image Analysis ### Overview The image displays a structured interface with three distinct panels, each containing an instruction, a visual element (image or diagram), and a corresponding response. The layout suggests a question-answering system that analyzes visual content to provide contextual explanations. ### Components/Axes 1. **Panel 1 (Top-Left)** - **Instruction**: "Which of the four subgraphs in the figure is taken at the same place as subgraph 1?" - **Visual**: A 2x2 grid of aerial photographs showing urban landscapes with varying densities of buildings, green spaces, and architectural features. - **Response**: - Analyzes urban density, building types, and circular structures (e.g., domes, spires). - Concludes that **Image 4** matches the urban density, building types, and circular structure of Image 1. 2. **Panel 2 (Center)** - **Instruction**: "What is the dome building in the picture?" - **Visual**: A cityscape featuring a tall, slender tower (CN Tower) and a large dome-shaped structure (Rogers Centre) in Toronto, Canada. - **Response**: - Identifies the dome as the **Rogers Centre**, a multi-purpose stadium with a retractable roof. - Notes its proximity to the CN Tower and its role in hosting sports events and conventions. 3. **Panel 3 (Bottom-Left)** - **Instruction**: "Where am I?" - **Visual**: A futuristic, neon-lit bar interior with holographic displays and cybernetic enhancements, resembling a scene from a video game. - **Response**: - Identifies the setting as **Night City** from *Cyberpunk 2077* (2077), a cyberpunk-themed environment with neon lighting and cybernetic characters. - Mentions mission/social interaction cues (e.g., "Sit next to Jackie"). ### Content Details - **Panel 1**: - Image 1: Dense urban area with a circular structure (possibly a dome/observatory). - Image 2: Suburban layout with large plots and fewer buildings. - Image 3: Urban layout with a church spire and spread-out design. - Image 4: Matches Image 1’s density, building types, and circular structure. - **Panel 2**: - Rogers Centre: Landmark in Toronto, recognizable by its retractable roof and proximity to the CN Tower. - **Panel 3**: - Cyberpunk 2077: Open-world RPG game set in a dystopian future (2077). - Key features: Neon lights, holographic displays, cybernetic enhancements. ### Key Observations 1. **Panel 1**: The response correctly identifies Image 4 as the match for Image 1 based on shared urban density and architectural features. 2. **Panel 2**: The Rogers Centre is accurately described, including its functional and geographical context. 3. **Panel 3**: The response aligns with the visual cues (neon, cybernetics) and contextual clues (game-specific references). ### Interpretation The interface demonstrates a system capable of analyzing visual content to infer location, architectural identity, and contextual narratives. The responses rely on: - **Spatial Analysis**: Comparing urban density, building types, and structural features (e.g., circular vs. spire). - **Cultural/Geographical Knowledge**: Recognizing landmarks like the Rogers Centre and CN Tower. - **Pop Culture References**: Linking the cyberpunk aesthetic to *Cyberpunk 2077*. The system’s accuracy suggests robust image recognition and contextual reasoning capabilities, though it lacks explicit uncertainty quantification (e.g., confidence scores for matches). The absence of numerical data or interactive elements implies a focus on descriptive analysis rather than quantitative modeling. </details> Figure 7: Kimi-VL exhibits strong visual reasoning capabilities by grounding visual content in spatial, contextual, and cultural knowledge. It accurately identifies matching urban locations based on structural and layout features, interprets scenes from video games like Cyberpunk 2077 using stylistic cues, and recognizes real-world landmarks such as the Rogers Centre in Toronto. 3.3 Reasoning Data Our reasoning data is meticulously constructed for activation and enhancement of the model’s multimodal reasoning capabilities during both the long-CoT supervised fine-tuning and reinforcement learning stages. Through developing a generation pipeline that resembles rejection sampling (RS) and prompt engineering, we collect and synthesize an amount of high-quality long-CoT data. Specifically, we first assemble a collection of QA data with ground truth annotations that require multi-step reasoning, such as mathematical problem-solving and domain-specific VQA. Subsequently, we sample multiple detailed reasoning trajectories for each question by leveraging a powerful long-CoT model - Kimi k1.5 \parencite team2025kimi with curated reasoning prompts. In rejection sampling, we feed the true labels and model predictions into an off-the-shelf reward model for judgment. Wrong chain-of-thought responses are filtered out according to the model evaluation as well as some rule-based rewards, thus improving the reasoning data quality. 4 Evaluation We begin by presenting our comprehensive model and conducting a comparative analysis with leading state-of-the-art (SoTA) solutions. Following this introduction, we proceed to assess various sub-capabilities of the model through detailed performance evaluations. This part examines how effectively the model handles different tasks and scenarios, providing insights into its strengths and limitations across diverse functional domains. 4.1 Comparison to the State-of-the-Art Models Table 3 presents a comprehensive evaluation of Kimi-VL against state-of-the-art vision-language models across multiple benchmarks. Although having a more parameter-efficient architecture (2.8B+0.4B activated parameters) compared to larger models such as GPT-4o, Llama-3.2-11B-Inst. and Gemma3-12B-IT, Kimi-VL demonstrates competitive or superior performance in several key areas. Our model employs a Mixture-of-Experts (MoE) architecture similar to DeepSeek-VL2, but outperforms it on most benchmarks with significantly fewer parameters (activated: 2.8B vs 4.5B; total: 16B vs 28B); it also outperforms Qwen2.5-VL-7B (actually 8.3B) on 19 out of 24 benchmarks, though the latter has 2.59 $×$ more activated parameters. The following sections analyze performance across specific domains, which reveals Kimi-VL ’s strengths in OCR, math, agent, long-form content understanding, multi-image and video perception. <details> <summary>x8.png Details</summary>  ### Visual Description ## Math Problem: Geometry Circle Theorem Application ### Overview The image contains a geometry problem involving a circle with points A, B, C, D, and center O. The problem asks for the measure of angle ACO given specific conditions about diameters and angles. A detailed solution is provided in Chinese, breaking down the problem into three logical steps. --- ### Components/Axes - **Diagram Elements**: - Circle with center **O** - Points labeled **A, B, C, D** on the circle - Line segments: **AB** (diameter), **AC, AD, BC, BD, CD** - Angles marked: **∠D = 62°**, **∠ACO** (unknown) - **Textual Elements**: - **Instruction Section** (Chinese): - Problem statement: "如图,AB为⊙O的直径,点D、C在⊙O上,∠D=62°,则∠ACO的度数为()" - Options: A. 26° B. 28° C. 30° D. 32° - **Response Section** (Chinese): - Step-by-step solution with mathematical reasoning --- ### Detailed Analysis #### 1. Problem Setup (Instruction Section) - **Key Facts**: - **AB** is the diameter of circle **O** ⇒ **∠ACB = 90°** (inscribed angle theorem) - **∠D = 62°** (inscribed angle subtended by arc **AC**) - Points **D** and **C** lie on the circle #### 2. Solution Steps (Response Section) 1. **Step 1: Inscribed Angle Theorem** - Since **AB** is a diameter, **∠ACB = 90°** (angle subtended by diameter) - **∠D = 62°** (given) 2. **Step 2: Central Angle Calculation** - Central angle **∠AOC** subtends the same arc **AC** as inscribed angle **∠D** - **∠AOC = 2 × ∠D = 2 × 62° = 124°** 3. **Step 3: Triangle Angle Sum** - In triangle **ACO**: - **∠OAC = ∠D = 62°** (inscribed angles subtending same arc) - **∠AOC = 124°** (from Step 2) - Sum of angles in triangle: **∠ACO + ∠OAC + ∠AOC = 180°** - **∠ACO + 62° + 124° = 180°** - **∠ACO = 180° - 186° = -8°** (Wait, this contradicts the solution. Correction: **∠OAC = ∠D = 62°** is incorrect. Re-evaluate.) - Correction: - **∠OAC = ∠D = 62°** is invalid. Instead, **∠OAC = ∠D = 62°** applies to a different arc. Re-examining the solution: - **∠OAC = ∠D = 62°** is incorrect. The correct relationship is **∠OAC = ∠D = 62°** only if they subtend the same arc. Rechecking the solution steps reveals a miscalculation. The correct derivation should be: - **∠OAC = ∠D = 62°** (inscribed angles subtending arc **OC**) - **∠AOC = 124°** (central angle) - **∠ACO = 180° - 62° - 124° = -6°** (still invalid). This indicates an error in the original solution. However, the final answer **28°** aligns with standard circle theorem applications, suggesting a transcription error in the intermediate steps. #### 3. Final Answer - **∠ACO = 28°** (Option B) --- ### Key Observations - The solution correctly applies the inscribed angle theorem (**∠ACB = 90°**) and central angle theorem (**∠AOC = 2 × ∠D**). - A minor inconsistency exists in the intermediate steps (angle sum calculation), but the final answer **28°** is mathematically valid. - The problem tests understanding of circle theorems and angle relationships in cyclic quadrilaterals. --- ### Interpretation This problem demonstrates the application of circle theorems to solve for unknown angles in a cyclic quadrilateral. The critical steps involve: 1. Recognizing **AB** as a diameter to establish a right angle (**∠ACB = 90°**). 2. Using the central angle theorem to relate **∠AOC** to **∠D**. 3. Applying the triangle angle sum property to isolate **∠ACO**. The final answer (**28°**) aligns with standard geometric principles, though the intermediate steps in the provided solution contain minor inconsistencies. This highlights the importance of verifying angle relationships in cyclic figures. </details> Figure 8: Kimi-VL demonstrates its capability to perform symbolic reasoning and geometric inference by solving a circle geometry problem step by step. The model analyzes given conditions, applies geometric theorems such as the inscribed angle theorem and properties of triangle angles, and accurately derives the target angle. <details> <summary>x9.png Details</summary>  ### Visual Description ## Financial Table: Sparkling Smiles Clinic Ratio Analysis ### Overview A financial performance table for Sparkling Smiles Clinic (SSC) showing key ratios across profitability, liquidity, efficiency, and activity metrics from 2004 to 2021. Data includes percentages and ratios for assets, liabilities, and operational metrics. ### Components/Axes - **Columns**: Years (2004–2021) - **Rows**: - **Profitability**: Net Profit Margin, Operating Profit Margin, ROA, ROE, EPS, P/E Ratio - **Liquidity**: Current Ratio, Quick Ratio, Cash Ratio, Working Capital - **Efficiency**: Inventory Turnover, Total Assets Turnover, Days Sales Outstanding, Days Payable Outstanding - **Activity**: Receivables Turnover, Payables Turnover, Inventory Days, Cash Conversion Cycle ### Detailed Analysis #### Profitability Ratios (2004–2021) - **Net Profit Margin**: 11% (2004) → 19% (2021) - **Operating Profit Margin**: 16% (2004) → 22% (2021) - **ROA**: 4% (2004) → 9% (2021) - **ROE**: 6% (2004) → 13% (2021) - **EPS**: $1.50 (2004) → $2.30 (2021) - **P/E Ratio**: 12x (2004) → 8x (2021) #### Liquidity Ratios (2004–2021) - **Current Ratio**: 1.8x (2004) → 2.5x (2021) - **Quick Ratio**: 1.3x (2004) → 1.9x (2021) - **Cash Ratio**: 0.9x (2004) → 1.5x (2021) - **Working Capital**: $450K (2004) → $1.2M (2021) #### Efficiency Ratios (2004–2021) - **Inventory Turnover**: 4.5x (2004) → 6.8x (2021) - **Total Assets Turnover**: 1.2x (2004) → 1.9x (2021) - **Days Sales Outstanding**: 45 days (2004) → 30 days (2021) - **Days Payable Outstanding**: 60 days (2004) → 40 days (2021) #### Activity Ratios (2004–2021) - **Receivables Turnover**: 8x (2004) → 12x (2021) - **Payables Turnover**: 5x (2004) → 7x (2021) - **Inventory Days**: 66 days (2004) → 53 days (2021) - **Cash Conversion Cycle**: 45 days (2004) → 30 days (2021) ### Key Observations - **Profitability**: All metrics improved steadily, with ROE nearly doubling. - **Liquidity**: Current and Quick Ratios increased, indicating stronger short-term financial health. - **Efficiency**: Inventory turnover rose 50%, while days outstanding/payable decreased by 30–50%. - **Activity**: Receivables turnover doubled, improving cash flow. ### Interpretation The clinic’s financial health improved significantly over 17 years, with stronger profitability, liquidity, and operational efficiency. The declining P/E ratio suggests reduced market expectations or stabilized growth. The cash conversion cycle shortened by 33%, reflecting better inventory and receivables management. --- ## Mathematical Formula: Bayesian Analysis ### Overview A Bayesian formula for updating posterior distributions using prior knowledge and likelihood functions. The formula includes terms for parameter estimation (α, β) and covariance matrices. ### Components/Axes - **Variables**: - **Data**: **x**_t-1, **x**₀ - **Parameters**: μ, Σ, α, β - **Covariance**: **I** (identity matrix) - **Structure**: - Posterior distribution **q** parameterized by **μ**, **Σ**. - Likelihood function **g** involves exponential terms with **α**, **β**. ### Detailed Analysis #### Formula Breakdown 1. **Posterior Distribution**: </details> Figure 9: Diverse OCR visualization. Kimi-VL demonstrates strong OCR capabilities across varied content types, including structured financial tables, complex mathematical formulas, and handwritten Chinese text. The model accurately parses tabular data into markdown, converts formulas to LaTeX, and transcribes handwritten paragraphs with contextual understanding, showcasing its versatility in multimodal text extraction and interpretation. <details> <summary>x10.png Details</summary>  ### Visual Description ## Screenshot: Chrome "Do Not Track" Feature Enablement Guide ### Overview The image shows a step-by-step guide to enable Chrome's "Do Not Track" feature, presented as a technical walkthrough with annotated screenshots. The guide is split into two columns (Steps 1-6 on the left, 7-12 on the right), each containing: - A browser window screenshot - A "Thought" section explaining the user's reasoning - An "Action" section describing the UI interaction - A "Toolcall" section with coordinates for automated interaction ### Components/Axes - **Header**: Blue box with instruction text: *"Can you enable the 'Do Not Track' feature in Chrome to enhance my online privacy?"* - **Main Content**: Two-column layout with 12 steps, each containing: - **Screenshot**: Browser window showing Chrome settings menu - **Thought**: Text explaining the user's decision-making process - **Action**: Description of UI interaction (e.g., "Click on the vertical ellipsis menu button") - **Toolcall**: Coordinates for programmatic interaction (e.g., `click(x=0.884, y=0.144)`) ### Detailed Analysis #### Steps 1-6 (Left Column) 1. **Step 1**: - **Screenshot**: Browser with Google homepage open - **Thought**: User identifies the need to access Chrome settings via the vertical ellipsis menu. - **Action**: Click the vertical ellipsis menu button. - **Toolcall**: `click(x=0.884, y=0.144)` 2. **Step 2**: - **Screenshot**: Expanded Chrome settings menu - **Thought**: User navigates to "Privacy and security" settings. - **Action**: Click "Privacy and security" option. - **Toolcall**: `click(x=0.365, y=0.317)` 3. **Step 3**: - **Screenshot**: Privacy and security settings page - **Thought**: User locates the "Do Not Track" option in the sidebar. - **Action**: Click "Privacy and security" sidebar menu. - **Toolcall**: `click(x=0.386, y=0.311)` 4. **Step 4**: - **Screenshot**: Privacy settings page with "Do Not Track" option - **Thought**: User scrolls to locate the "Do Not Track" setting. - **Action**: Scroll down to reveal more options. - **Toolcall**: `scroll(-5)` 5. **Step 5**: - **Screenshot**: Privacy settings page with "Do Not Track" toggle - **Thought**: User identifies the "Do Not Track" toggle switch. - **Action**: Click the toggle switch. - **Toolcall**: `click(x=0.527, y=0.884)` 6. **Step 6**: - **Screenshot**: Privacy settings page with "Manage certificates" option - **Thought**: User navigates to certificate management settings. - **Action**: Click "Manage HTTPS/SSL certificates and settings." - **Toolcall**: `click(x=0.527, y=0.883)` #### Steps 7-12 (Right Column) 7. **Step 7**: - **Screenshot**: "Manage certificates" page - **Thought**: User realizes they need to return to the previous page. - **Action**: Click the back button. - **Toolcall**: `click(x=0.486, y=0.242)` 8. **Step 8**: - **Screenshot**: Privacy and security settings page - **Thought**: User selects "Security" option for detailed controls. - **Action**: Click "Security" option. - **Toolcall**: `click(x=0.522, y=0.884)` 9. **Step 9**: - **Screenshot**: Security settings page - **Thought**: User returns to the previous page to continue navigation. - **Action**: Click the back button. - **Toolcall**: `click(x=0.486, y=0.242)` 10. **Step 10**: - **Screenshot**: Privacy and security settings page - **Thought**: User selects "Third-party cookies" option. - **Action**: Click "Third-party cookies" option. - **Toolcall**: `click(x=0.542, y=0.772)` 11. **Step 11**: - **Screenshot**: "Send a 'Do Not Track' request" dialog - **Thought**: User enables the "Send a 'Do Not Track' request" toggle. - **Action**: Click the toggle switch. - **Toolcall**: `click(x=0.808, y=0.871)` 12. **Step 12**: - **Screenshot**: Confirmation dialog for "Do Not Track" - **Thought**: User confirms the action to enable the feature. - **Action**: Click "Confirm" button. - **Toolcall**: `click(x=0.708, y=0.669)` ### Key Observations - The guide emphasizes navigating Chrome's settings hierarchy to locate privacy controls. - Coordinates in toolcalls suggest a focus on programmatic automation (e.g., `x=0.884, y=0.144` for the vertical ellipsis menu). - The final confirmation dialog (Step 12) validates successful enablement of the feature. ### Interpretation This guide demonstrates a systematic approach to configuring Chrome's privacy settings, highlighting the importance of: 1. **Hierarchical Navigation**: Users must traverse multiple menus (Settings → Privacy → Security) to access granular controls. 2. **Automation Readiness**: The inclusion of precise coordinates implies this guide may be used for automated testing or scripting. 3. **Privacy Trade-offs**: The "Do Not Track" feature, while intended to enhance privacy, relies on website cooperation, as noted in the absence of explicit confirmation in the final step. The structured breakdown of UI interactions provides a clear roadmap for both manual and automated configuration of browser privacy settings. </details> Figure 10: Kimi-VL is capable of following multi-step reasoning processes to complete complex GUI tasks. In this example, it successfully enables the “Do Not Track” feature in the Chrome browser to enhance online privacy. The agent interprets each screen, identifies relevant UI elements, and performs the appropriate actions sequentially with clear thoughts, actions, and API calls. <details> <summary>x11.png Details</summary>  ### Visual Description ## Screenshot: Technical Document with Scene Analysis ### Overview The image shows a two-column document. The left column contains a grid of small thumbnail images (likely video frames) with a header labeled "Instruction." The right column contains a text block titled "Response" with detailed scene descriptions. The document appears to analyze a video by splitting it into timestamped scenes with narrative descriptions. ### Components/Axes - **Left Column**: - Header: "Instruction" (bold text) - Grid: 12 rows × 12 columns of thumbnail images (total 144 frames) - Thumbnails vary in content (e.g., dark scenes, snowy landscapes, human faces, religious symbols) - **Right Column**: - Header: "Response" (bold text) - Text: 12 paragraphs, each starting with a timestamp (e.g., `[00:00:00 - 00:00:15]`) followed by detailed scene descriptions ### Detailed Analysis #### Left Column ("Instruction") - **Grid Structure**: - 12 rows × 12 columns (144 total thumbnails) - Thumbnails are small (≈100x100px) and densely packed - Content varies significantly: - Dark/black frames (possibly video start/end) - Snowy mountain landscapes - Human faces (close-ups) - Religious symbols (prayer wheels, crosses) - Natural scenes (forests, waterfalls) - Architectural elements (buildings, bridges) #### Right Column ("Response") - **Text Structure**: - 12 paragraphs, each corresponding to a video scene - Each paragraph follows this format: 1. Timestamp (e.g., `[00:00:00 - 00:00:15]`) 2. Detailed scene description (≈150-200 words) 3. Narrative elements (e.g., "a person cooking food," "snow-covered mountain range") - **Key Descriptions**: 1. **Scene 1** (`[00:00:00 - 00:00:15]`): Dark room with a single light source, person cooking, mysterious atmosphere 2. **Scene 2** (`[00:00:15 - 00:00:36]`): Elderly person's face, furred brow, prayer wheel, spiritual reflection 3. **Scene 3** (`[00:00:36 - 00:00:46]`): Snow-capped mountains, aerial view, natural grandeur 4. **Scene 4** (`[00:00:46 - 00:01:09]`): Prayer wheel close-up, dynamic motion, spiritual symbolism 5. **Scene 5** (`[00:01:09 - 00:01:27]`): Prayer wheel in motion, blurred background, spiritual focus 6. **Scene 6** (`[00:01:27 - 00:01:39]`): Yellow jacket, snowy landscape, human-nature interaction 7. **Scene 7** (`[00:01:39 - 00:02:03]`): Snow-covered mountains, treacherous terrain, spiritual journey 8. **Scene 8** (`[00:02:03 - 00:02:17]`): Prayer wheel close-up, symbolic light, spiritual contemplation 9. **Scene 9** (`[00:02:17 - 00:02:42]`): Skiing in snowstorm, human-nature conflict, survival themes 10. **Scene 10** (`[00:02:42 - 00:02:46]`): Prayer wheel close-up, spiritual symbolism, cultural context 11. **Scene 11** (`[00:03:05 - 00:03:27]`): Forest at sunset, spiritual atmosphere, human-nature connection 12. **Scene 12** (`[00:03:27 - 00:03:37]`): Dark cave interior, mysterious mood, symbolic exploration ### Key Observations 1. **Temporal Progression**: Scenes transition from indoor (cooking) to outdoor (mountains) to spiritual (prayer wheels) themes 2. **Symbolism**: Prayer wheels appear in 4 scenes, suggesting recurring spiritual motifs 3. **Natural Elements**: Snow, mountains, and forests dominate 60% of descriptions 4. **Human Presence**: Humans appear in 8 scenes, often in contemplative or action-oriented poses 5. **Lighting**: Contrast between dark indoor scenes and bright outdoor/nature scenes ### Interpretation This document appears to be a technical analysis of a video, breaking it into discrete scenes with precise timestamps and narrative descriptions. The recurring spiritual elements (prayer wheels, religious symbols) suggest the video may document a religious or cultural journey. The progression from mundane (cooking) to sublime (mountain landscapes) to symbolic (prayer wheels) indicates a narrative arc of human experience and spiritual reflection. The detailed descriptions emphasize sensory elements (light, texture, sound) to create immersive scene transitions, typical of documentary or artistic video production. </details> Figure 11: Video scene splitting. Kimi-VL processes a long-form video by segmenting it into coherent scenes and providing detailed start/end timestamps along with fine-grained natural language descriptions for each scene. footnotemark: <details> <summary>x12.png Details</summary>  ### Visual Description ## Screenshot: Presentation Slide ### Overview The image shows a presentation slide with two main sections: 1. **Instruction**: A Chinese proverb about teaching skills, followed by a task to explain a "further requirement." 2. **Response**: A textual answer interpreting the proverb, emphasizing motivation and lifelong learning. ### Components/Axes - **Instruction Section**: - Text: "There is an old Chinese saying: Give a man a fish, and you feed him for a day; teach a man to fish, and you feed him for a lifetime. In this representation, the author puts forward a further requirement. Find it then explain it in detail." - Visual Elements: - 10 sequential images of a presenter (likely a video playback interface) with timestamps (00:00 to 35:55). - A "Thank you!" slide with a Twitter handle (`@hwochung27`). - **Response Section**: - Text: "The requirement is to 'Teach him the taste of fish and make him hungry.' This implies that in addition to teaching someone how to perform a task or acquire a skill, it is also important to inspire and motivate them to continue learning and improving on their own. By making them hungry for more knowledge or experience, they will be more likely to seek out new challenges and opportunities for growth, which can lead to greater success and fulfillment in the long run." ### Detailed Analysis - **Instruction Text**: - The proverb is presented as a starting point for the task. - The task explicitly asks to identify and explain a "further requirement" beyond the original proverb. - **Response Text**: - The answer interprets the "further requirement" as teaching the "taste of fish" (metaphor for intrinsic motivation) and making the learner "hungry" (metaphor for sustained curiosity). - Emphasizes the importance of fostering self-driven learning beyond basic skill acquisition. - **Visual Elements**: - The 10 images show a presenter in a classroom setting, likely delivering the content. - The final slide includes a "Thank you!" message and a Twitter handle, suggesting a professional or educational context. ### Key Observations 1. The slide structure mirrors the proverb’s logic: a problem (instruction) followed by a solution (response). 2. The response directly addresses the task by rephrasing the proverb’s "teach a man to fish" into a modern educational framework. 3. The Twitter handle implies the content may be part of a broader discussion or series. ### Interpretation The slide illustrates a pedagogical philosophy that extends the original proverb: teaching skills is insufficient without cultivating intrinsic motivation. The "taste of fish" metaphor suggests that learners must not only gain competence but also develop a desire for continuous growth. This aligns with theories of self-determination in education, where autonomy and purpose drive long-term engagement. The response’s focus on "hunger for knowledge" implies that extrinsic rewards (e.g., feeding) are less impactful than internal drive. The inclusion of a Twitter handle hints at a broader discourse on education or personal development. **Note**: No numerical data, charts, or diagrams are present. The analysis is based solely on textual content and visual context. </details> Figure 12: Catching and understanding key details from an hour-long video course. Kimi-VL demonstrates its ability to comprehend and interpret instructional video content by analyzing frame sequences and extracting conceptual progression over time. In this case, the model identifies a deepening of the traditional saying “Teach a man to fish, and you feed him for a lifetime” into a more nuanced idea: “Teach him the taste of fish and make him hungry.” footnotemark: 4.1.1 College-level Academic Problems Our Kimi-VL model demonstrates competitive performance on college-level academic benchmarks. On MMMU validation set, it achieves a score of 57.0%, which outperforms DeepSeek-VL2 (51.1%) and is comparable to Qwen2.5-VL-7B (58.6%) and even Gemma-3-12B-IT (59.6%), despite having significantly fewer activated parameters. On video college-level problems, it significantly outperforms Qwen2.5-VL-7B and DeepSeek-VL2, only behind >10B Gemma-3-12B-IT, demonstrating reasonable university-level understanding capabilities compared to larger models. These results indicate that Kimi-VL effectively balances parameter efficiency with academic reasoning abilities. 4.1.2 General Visual Ability Kimi-VL exhibits strong general visual understanding capabilities across multiple benchmarks. On MMBench-EN-v1.1, it achieves 83.1% accuracy, outperforming all efficient VLMs in comparison, and performing on par with GPT-4o. For AI2D, our model achieves 84.9% and surpasses all compared models including GPT-4o (84.6%). On MMVet, Kimi-VL scores 66.7% and ties closely with Qwen2.5-VL-7B (67.1%) and GPT-4o-mini (66.9%). For RealWorldQA, it achieves 68.1%, outperforming Gemma3-12B (59.1%) and approaching Qwen2.5-VL-7B (68.5%). These results demonstrate that our model maintains robust general visual understanding despite its compact architecture. In multi-image reasoning tasks, Kimi-VL shows promising capabilities with a score of 57.3% on the BLINK benchmark. This performance surpasses Qwen2.5-VL-7B (56.4%), GPT-4o-mini (53.6%), Gemma3-12B-IT (50.3%), and Llama3.2-11B-Inst. (39.8%). The ability to reason across multiple images requires understanding spatial and temporal relationships between visual elements, which our model handles effectively with fewer parameters than most competitors. 4.1.3 Mathematical Reasoning With its relatively small scale, Kimi-VL also demonstrates strong mathematical reasoning capabilities, particularly on the MathVista benchmark where it achieves 68.7%, outperforming all compared models including GPT-4o (63.8%) and Qwen2.5-VL-7B (68.2%). It indicates our model’s exceptional ability to understand and solve mathematical problems presented in visual contexts. On the more challenging MathVision benchmark, due to limited activated parameters, Kimi-VL outperforms DeepSeek-VL2 and Llama-3.2-11B-Inst., but lags behind Qwen2.5-VL-7B and Gemma-12B-IT. Nevertheless, through RL and test-time scaling, Kimi-VL-Thinking has significantly improved and already on par with 30B-level VLMs (see Table 4). These results highlight our model’s effectiveness in combining visual perception with mathematical problem-solving, an essential capability for real-world applications. 4.1.4 Document Understanding and OCR Kimi-VL excels in document understanding and OCR tasks across all benchmarks in this category. On InfoVQA, it achieves 83.2% accuracy, outperforming GPT-4o (80.7%) and DeepSeek-VL2 (78.1%). For OCRBench, our model scores 86.7%, surpassing all other models including GPT-4o-mini (78.5%) and DeepSeek-VL2 (81.1%). These results demonstrate that our model has exceptional text recognition and document understanding capabilities, making it especially suitable for applications involving document processing and information extraction. 4.1.5 Agent Grounding and Multi-turn Agent Interaction In agent-based tasks, Kimi-VL demonstrates remarkable performance. On single-step grounding, our model shows strong accuracy, with 92.0% on ScreenSpot-V2 and 34.5% on extremely difficult ScreenSpot-Pro (on 4K screens), proving its strong agent grounding abilities. More importantly, it also shows strong multi-step turn agent interaction abilities: For OSWorld, Kimi-VL reaches 8.22%, outperforming GPT-4o (5.03%) and other capable open-source models; On WindowsAgentArena, our model achieves 10.4%, also surpassing GPT-4o (9.4%) and others. These results highlight Kimi-VL’s exceptional ability to understand and interact with operating system interfaces, suggesting strong potential for applications in automated UI navigation and task execution. 4.1.6 Long Document and Long Video Understanding Kimi-VL demonstrates competitive performance in long-form content understanding. On MMLongBench-Doc, a challenging benchmark with question-answering on up to 100+ pages, it achieves 35.1%, outperforming GPT-4o-mini (29.0%) and Qwen2.5-VL-7B (29.6%), only behind GPT-4o (42.8%). For long video understanding, on Video-MME, our model outperforms all efficient VLMs and especially leads on the fairer w/o subtitle setting, where models have to find answers from video frames instead of hacking from input subtitles; on w/ subtitle setting, it also reaches extraordinary 72.6% accuracy. On the MCQ subset of MLVU, Kimi-VL achieves an impressive 74.2% score, achieving state-of-the-art and surpassing both GPT-4o (64.6%) and Qwen2.5-VL-7B (70.2%). For LongVideoBench, it scores 64.5%, outperforming all compared models except GPT-4o (66.7%). These results demonstrate Kimi-VL ’s strong capability to understand long-form PDFs and videos. 4.1.7 Egocentric and Fine-grained Video Perception Kimi-VL also shows strong performance in more nuanced video perception tasks. On EgoSchema full set (hidden test set), it achieves 78.5%, significantly outperforming GPT-4o (72.2%), Qwen2.5-VL-7B (65.0%). For VSI-Bench, a very challenging benchmark that requires to understand spatial relationships and correspondences of multiple objects in a video, our model scores 37.4%, surpassing GPT-4o (34.0%) and Qwen2.5-VL-7B (34.2%). In TOMATO that examines fine-grained temporal perception of VLMs, Kimi-VL reaches 31.7%, outperforming Qwen2.5-VL-7B (27.6%) and GPT-4o-Mini (28.8%). These results demonstrate our model’s strong capability to understand dynamic visual content, track objects over time, and interpret complex actions in video sequences, making it well-suited for applications requiring temporal visual understanding. 4.2 Kimi-VL-A3B-Thinking: A Reasoning Extension of Kimi-VL Furthermore, we conduct a reasoning extension to empower Kimi-VL to reason with CoT and present a long-thinking version of the model, Kimi-VL-Thinking, through long-CoT activation and reinforcement learning. We validate its superior performance on several image benchmarks, as shown in Table 4. | MathVision (full) (Pass@1) MathVista (mini) (Pass@1) MMMU (val) (Pass@1) | 30.4 63.8 69.1 | - 56.7 60.0 | 38.1 74.8 74.8 | 25.1 68.2 58.6 | 35.5 62.3 64.8 | 32.1 56.4 59.6 | - 71.0 77.3 | 35.9 71.4 70.3 | 38.6 74.9 70.0 | 36.8 71.3 61.7 | 56.9 80.1 64.0 | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | MMMU-Pro (avg) (Pass@1) | 51.7 | 37.6 | 51.1 | 38.1 | - | 32.1 | - | - | - | 43.0 | 46.3 | | VideoMMMU (Pass@1) | 61.1 | - | 60.2 | 47.0 | 61.8 | 57.2 | - | - | - | 55.5 | 65.2 | Table 4: Performance of Kimi-VL-Thinking and Kimi-VL-Thinking-2506 on multimodal reasoning benchmarks. The metrics evaluated include MathVista (mini), MMMU (val), MMMU-Pro (average), MathVision (full) and VideoMMMU, with results expressed in Pass@1. The Kimi-VL-Thinking-2506 performs well in most cases, showcasing the enhanced reasoning and processing capabilities of the “thinking” variant across different domains and scales. <details> <summary>x13.png Details</summary>  ### Visual Description ## Scatter Plots: Test Time Accuracy vs. Max Thinking Length ### Overview The image contains three scatter plots comparing **Test Time Accuracy (%)** against **Max Thinking Length (k tokens)** for three different systems: **MathVision**, **MathVista**, and **MMU**. Each plot shows a clear upward trend, indicating that longer thinking lengths correlate with higher accuracy. --- ### Components/Axes 1. **X-Axis (Horizontal)**: - Label: **Max Thinking Length (k tokens)** - Values: 1, 2, 4, 8, 16 (logarithmic scale). 2. **Y-Axis (Vertical)**: - Label: **Test Time Accuracy (%)** - Ranges: - MathVision: 16%–36% - MathVista: 66%–71% - MMU: 48%–62% 3. **Legends**: - Positioned at the top of each plot, labeled with the system name (MathVision, MathVista, MMU). - Data points are black dots with percentage labels. --- ### Detailed Analysis #### MathVision - **Data Points**: - 1k tokens: 18.7% - 2k tokens: 22.6% - 4k tokens: 29.0% - 8k tokens: 34.0% - 16k tokens: 36.8% - **Trend**: Gradual increase in accuracy with longer thinking lengths. #### MathVista - **Data Points**: - 1k tokens: 66.7% - 2k tokens: 69.0% - 4k tokens: 70.6% - 8k tokens: 70.9% - 16k tokens: 71.3% - **Trend**: Steeper initial improvement, plateauing near 71% at 16k tokens. #### MMU - **Data Points**: - 1k tokens: 49.2% - 2k tokens: 52.4% - 4k tokens: 56.2% - 8k tokens: 60.1% - 16k tokens: 61.7% - **Trend**: Consistent upward trajectory, but slower growth compared to MathVista. --- ### Key Observations 1. **MathVista** achieves the highest accuracy (71.3% at 16k tokens), outperforming both MathVision and MMU. 2. **MathVision** shows the lowest baseline accuracy (18.7% at 1k tokens) but improves significantly with longer thinking lengths. 3. **MMU** has moderate accuracy (61.7% at 16k tokens) with a steady but gradual improvement. 4. All systems exhibit diminishing returns at higher thinking lengths (e.g., MathVista’s accuracy increases by only 0.4% between 8k and 16k tokens). --- ### Interpretation - **Performance Correlation**: Longer thinking lengths consistently improve accuracy across all systems, suggesting that extended computation time allows for better problem-solving. - **Model Efficiency**: - **MathVista** demonstrates the most efficient scaling, achieving near-peak accuracy (71.3%) with minimal additional gains at 16k tokens. - **MathVision** requires the longest thinking length (16k tokens) to reach its peak (36.8%), indicating potential inefficiencies in its reasoning process. - **MMU** balances moderate accuracy with steady improvement, though it lags behind MathVista in both baseline and peak performance. - **Diminishing Returns**: The flattening trends at higher thinking lengths (e.g., MathVista’s 70.9% → 71.3% increase) suggest a practical limit to the benefits of extended computation. This data highlights the trade-off between computational resources and accuracy, with MathVista emerging as the most effective system for maximizing test time accuracy. </details> Figure 13: Test-time accuracy when scaling the max thinking token length of our Kimi-VL-Thinking model. Kimi-VL-Thinking significantly improves over the base Kimi-VL model, with gains of 2.6% on MathVista, 4.7% on MMMU, and 15.4% on MathVision, demonstrating its capability to leverage test-time computation for deeper reasoning and better handling of complex multimodal queries. In Table 4, Kimi-VL-Thinking further outperforms or rivals state-of-the-art thinking and non-thinking models: achieving 71.3% on MathVista, outperforming GPT-4o (63.8%) and GPT-4o-mini (56.7%); scoring 61.7% on MMMU, surpassing GPT-4o-mini (60.0%) and Qwen2.5-VL-7B (58.6%); and reaching 36.8% on MathVision, exceeding GPT-4o (30.4%) and Gemma-3-27B-IT (35.5%), even QVQ-72B (35.9%). While marginally behind some larger-scale models on select benchmarks, Kimi-VL-Thinking accomplishes these results with only 3B activated parameters—orders of magnitude fewer than its counterparts—underscoring its strong efficiency and effectiveness in multimodal reasoning. Our Kimi-VL-Thinking model also exhibits strong test-time scaling properties, as shown in Figure 13. Specifically, increasing the max thinking token length at inference time consistently improves test-time accuracy across all three benchmarks. For example, on MathVision, the accuracy rises steadily from 18.7% at 1k tokens to 36.8% at 16k tokens, and similar upward trend is also observed on MMMU, indicating that the model is able to utilize longer reasoning chains for better performance. However, not all benchmarks benefit equally from longer thinking lengths. On MathVista, performance saturates early, with accuracy reaching 70.9% at 4k tokens and no further significant gains observed as the token length increases to 16k. It suggests that for this task, the necessary reasoning depth is already captured within a relatively short context, and additional computation does not yield further improvements. 4.3 Kimi-VL-A3B-Thinking-2506: From Reasoning Extension to Integrated Thinking Model Table 5: Performance of Kimi-VL-A3B-Thinking-2506 on multimodal benchmarks that do not require extensive reasoning. | Benchmark (Metric) General Multimodal MMBench-EN-v1.1 (Acc) | GPT-4o 83.1 | Qwen2.5- VL-7B 83.2 | Gemma3- 12B-IT 74.6 | Kimi-VL-A3B- Instruct 82.9 | Kimi-VL-A3B- Thinking 76.0 | Kimi-VL-A3B- Thinking-2506 84.4 | | --- | --- | --- | --- | --- | --- | --- | | RealWorldQA (Acc) | 75.4 | 68.5 | 59.1 | 68.1 | 64.0 | 70.0 | | OCRBench (Acc) | 815 | 864 | 702 | 864 | 864 | 869 | | MMStar (Acc) | 64.0 | 63.0 | 56.1 | 61.7 | 64.2 | 70.4 | | MMVet (Acc) | 69.1 | 67.1 | 64.9 | 66.7 | 69.5 | 78.1 | | Video | | | | | | | | MMVU ${}_{\text{val}}$ (Pass@1) | 67.4 | 50.1 | 57.0 | 52.7 | 53.0 | 57.5 | | Video-MME (w/ sub.) (Acc) | 77.2 | 71.6 | 62.1 | 72.7 | 66.0 | 71.9 | | OS-Agent Grounding | | | | | | | | ScreenSpot-Pro (Acc) | 0.8 | 29.0 | — | 35.4 | — | 52.8 | | ScreenSpot-V2 (Acc) | 18.1 | 84.2 | — | 92.8 | — | 91.4 | | OSWorld-G (Acc) | - | 31.5 | — | 41.6 | — | 52.5 | | Long Document | | | | | | | | MMLongBench-Doc (Acc) | 42.8 | 29.6 | 21.3 | 35.1 | 32.5 | 42.1 | While Kimi-VL-A3B-Thinking shows excellent thinking abilities on hard reasoning tasks, we further provide the updated Kimi-VL-A3B-Thinking-2506 Tech Blog: https://huggingface.co/blog/moonshotai/kimi-vl-a3b-thinking-2506, a new reasoning variant that is not only smarter, but integrates key abilities of Kimi-VL-A3B-Instruct (perception, video, long-document, and OS-agent abilities) into this thinking model. Kimi-VL-Thinking-2506 significantly improves reasoning efficiency while reducing token consumption. As shown in Table 4, Kimi-VL-Thinking-2506 achieves 56.9% on MathVision (+20.1% improvement on original Kimi-VL-Thinking), 80.1% on MathVista (+8.4%), 46.3% on MMMU-Pro (+3.2%), and 64.0% on MMMU (+2.1%), demonstrating non-trivial gains across multiple reasoning benchmarks. Meanwhile, while solving these hard reasoning problems, the 2506 version reduces the average output token length by around 20% (e.g., 2.9K $→$ 2.4K on MMMU-val and 5.8K $→$ 4.4K on MathVision), facilitating it to be more efficient and user-friendly for practical deployments. Beyond extensive reasoning tasks, Kimi-VL-Thinking demonstrates stronger visual perception capabilities (Table 5). Compared to the previous non-thinking variant (Kimi-VL-A3B-Instruct), Kimi-VL-A3B-Thinking-2506 achieves competitive or superior results on general multimodal understanding benchmarks: 84.4% on MMBench-EN-v1.1, 70.4% on MMStar, 70.0% on RealWorldQA, and 78.4% on MMVet, underscoring its broader competence in vision-language tasks. In terms of token efficiency, the 2506 version only requires in average 180 tokens per answer when solving MMBench, 1/3 compared to the previous thinking model while improving 8.4% accuracy. Kimi-VL-A3B-Thinking-2506 also extends its reasoning ability to video and long-context domains. It establishes new state-of-the-art results among open-source models on VideoMMMU (65.2%, 4% better than GPT-4o), a challenging video reasoning benchmark; it also maintains robust general video understanding performance with 71.9% on Video-MME, matching the long video understanding ability of Kimi-VL-A3B-Instruct. It also scores 42.1% (first open-source model matching GPT-4o) on MMLongBench-Doc (Table 5), a 10% improvement over the previous thinking model and 7% over the previous instruct model, demonstrating its robust ability on broader long-form visual inputs. As mentioned in the method part, the continual training on MoonViT (3.2 million max input pixels) of Kimi-VL-A3B-Thinking-2506 leads to substantial improvements on high-resolution perception and OS grounding benchmarks, achieving 83.2% on V* Benchmark (without external tools), 52.8% on ScreenSpot-Pro, and 52.5% on OSWorld-G (full set with refusal samples), showing huge improvements over both previous variants. We hope that this high-resolution multimodal reasoning model brings about interesting new capabilities in the real world. 5 Conclusion, Limitation, and Future Work We introduce Kimi-VL, a VLM designed with a balanced approach to cover both multimodal and text-only pre-training/post-training, underpinned by an MoE-based architecture for scalable efficiency. Its 128K extended context window enables precise retrieval in lengthy texts and videos, while the native-resolution encoder MoonViT helps maintain high accuracy with low computational overhead in ultra-high-resolution visual tasks. Additionally, Kimi-VL-Thinking facilitates effective long-chain reasoning in complex image and video inference. Overall, Kimi-VL demonstrates robust adaptability and efficiency across multimodal, long-context, and high-resolution tasks, indicating substantial potential for future research and industrial applications. However, Kimi-VL still faces several challenges: 1. Although the current model size performs effectively for many standard tasks, it remains too limited to address highly specialized or domain-specific problems, or problems that are strongly dependent on language abilities, restricting Kimi-VL’s ability to handle extremely complex scenarios. 1. While the reasoning capability is already strong for typical use cases, it has yet to reach its theoretical upper bound, particularly for intricate tasks requiring multi-step inference or deeper contextual understanding. 1. Despite providing a 128K extended context window, due to limited parameters in its attention layers (which is only comparable to a 3B model), its long-context abilities is still insufficient for certain advanced applications that involve extremely long sequences or high-volume contextual information. In the future, we will tackle these challenges by scaling up the model size, expanding pre-training data, and enhancing post-training algorithms. Our next steps include optimizing Kimi-VL and releasing larger versions, as well as refining post-training and test-time scaling mechanisms for a better thinking model. These efforts will pave the way for more advanced applications in both research and industry. \printbibliography [title=References] Appendix Appendix A Contributions Core Contributors Bohong Yin Bowei Xing Cheng Chen Chu Wei Dehao Zhang Dongliang Wang Haoning Wu ∗ Haotian Yao Haoyu Lu ∗ Hao Yang Kun Ouyang Lin Sui Xinyuan Wang # Xinyu Zhou Yang Li Y. Charles ∗ Yiping Bao Yimin Chen Yuanxin Liu Yuxin Wu Zaida Zhou Zhaowei Li Zhiqi Huang Zhilin Yang Ziwei Chen Contributors Angang Du Bowen Qu Bowen Wang # Chenlin Zhang Chenzhuang Du Congcong Wang Dikang Du Enming Yuan Enzhe Lu Fang Li Flood Sung Guangda Wei Guokun Lai Han Zhu Hao Ding Hao Hu Hao Zhang Heng Wang Hongcheng Gao Huabin Zheng Jiaming Li Jianlin Su Jianzhou Wang Jiaqi Deng # Jiezhong Qiu Jin Xie Jinhong Wang Jingyuan Liu Junjie Yan Liang Chen Longhui Yu Mengfan Dong Mengnan Dong Nuo Xu Pengyu Cheng Qizheng Gu Runjie Zhou Shaowei Liu Sihan Cao Tao Yu # Tianhui Song Tongtong Bai Weiran He Wei Song Weixiao Huang Weixin Xu Xiaokun Yuan Xingzhe Wu Xingcheng Yao Xinhao Li Xinxing Zu Yangyang Hu Yan Zhong Yanru Chen Yibo Miao Yejie Wang Yibo Liu Yidao Qin Yiqin Wang Yongsheng Kang Yuhao Dong Yulun Du Yuzhi Wang Yuzi Yan Zhejun Jiang Zheng Zhang Zihao Huang Zijia Zhao Zongyu Lin * Project lead(s). # The University of Hong Kong, Moonshot.ai The listing of authors is in alphabetical order based on their first names. Appendix B Evaluation Details B.1 Image Benchmark MMMU \parencite yue2024mmmu encompasses a carefully curated collection of 11.5K multimodal questions sourced from college exams, quizzes, and textbooks. These questions span six major academic fields: Art & Design, Business, Science, Health & Medicine, Humanities & Social Science, and Tech & Engineering. MMBench-EN-v1.1 \parencite MMBench is a fine-grained benchmark that contains 2974 multiple-choice questions, covering 20 ability dimensions. It incorporate perception and reasoning as the top-level ability dimensions in its ability taxonomy, leading to different levels of evaluation in various ability dimensions. MMStar \parencite chen2024mmstar is an elite vision-indispensable multimodal benchmark comprising 1,500 challenge samples meticulously selected by humans. It is designed to benchmark 6 core capabilities and 18 detailed axes, aiming to evaluate the multimodal capacities of LVLMs with a carefully balanced and purified selection of samples. MMVet \parencite yu2024mmvet is designed based on the insight that the intriguing ability to solve complicated tasks is often achieved by a generalist model being able to integrate different core vision-language capabilities. It defines 6 core VL capabilities and examines the 16 integrations of interest derived from the capability combination. RealWorldQA \parencite realworldQA is a benchmark designed to evaluate the real-world spatial understanding capabilities of multimodal models. It assesses how well the models comprehend physical environments. The benchmark consists of over 700 images, each accompanied by a question and a verifiable answer, and these images are drawn from various real-world scenarios. AI2D \parencite kembhavi2016ai2d is a dataset of over 5000 grade school science diagrams with over 150000 rich annotations, their ground truth syntactic parses, and more than 15000 corresponding multiple choice questions. MathVision \parencite wang2024measuring is a carefully curated collection of 3,040 high-quality mathematical problems with visual contexts that are sourced from real math competitions. It covers 16 distinct mathematical disciplines and is graded across 5 levels of difficulty. This dataset offers a comprehensive and diverse set of challenges, making it ideal for evaluating the mathematical reasoning abilities of LMMs. MathVista \parencite lu2023mathvista is a benchmark that integrates challenges from a variety of mathematical and visual tasks, demanding participants to exhibit fine-grained, deep visual understanding along with compositional reasoning to successfully complete the tasks. BLINK \parencite fu2024blink is a benchmark designed to evaluate multi-image visual cognition, encompassing tasks related to depth relationships, feature matching, digital forensics, and spatiotemporal reasoning. It features a diverse set of multi-image perceptual similarity tasks, validated through standardized protocols. InfoVQA \parencite mathew2022infographicvqa is a dataset specifically designed to assess models’ capabilities in interpreting and reasoning with complex infographics that integrate text, graphics, and visual elements. Model performance on this dataset is evaluated using the ANLS metric on the test set. OCRBench \parencite liu2023hidden evaluates the OCR capabilities of MLLMs across five tasks: text recognition, scene text VQA, document VQA, key information extraction, and handwritten math expression recognition. The benchmark is scored out of a maximum of 1000 points. B.2 Video and Long Document Benchmark VideoMMMU \parencite arxiv2025videommmu is a video benchmark designed to evaluate the college-level knowledge acquisition capabilities of large multimodal models. It consists of 300 expert-level videos and 900 human-annotated questions. The videos span six diverse academic disciplines: Art, Humanities, Medicine, Business, Science, and Engineering. The questions are structured to align with three cognitive stages: Perception, Comprehension, and Adaptation. MMVU \parencite arxiv2025mmvu is a video benchmark designed to evaluate the expert-level video understanding ability. The benchmark contains 3,000 expert-annotated questions over 1,529 videos, which span 27 subjects from four core disciplines: Science, Healthcare, Humanities & Social Sciences, and Engineering. Video-MME \parencite arxiv2024videomme is a video benchmark that consists of 900 manually selected videos (totaling 254 hours length), and 2,700 QA pairs. The videos, varying in duration, are categorized into 30 fine-grained classes across six diverse domains: Knowledge, Film & Television, Sports Competition, Artistic Performance, Life Record, and Multilingual content. Evaluations are conducted under two different settings: with and without subtitles. MLVU \parencite arxiv2024mlvu is designed to evaluate the model performance in comprehending long videos from multiple aspects. It consists of 1,730 videos along with 3,102 corresponding question-answer pairs (2,593 in dev set and 509 in test set). Videos of this benchmark are collected from multiple scenarios, including Sport, Ego-centric, Life Record, Tutorial, etc. The close-ended task set of MLVU comprises 7 different tasks: Action Order, Action Count, Topic Reasoning, Anomaly Recognition, Plot QA, Ego Reasoning, and Needle QA. LongVideoBench \parencite nips2024longvideobench is a video question-answering benchmark designed to evaluate the long-form multimodal perception and relation capability of large multimodal models. The benchmark includes 3,763 web-collected videos spanning various lengths and themes, along with their corresponding subtitles. It includes 6,678 human-annotated multiple-choice questions, distributed across 17 fine-grained categories, which accesses different aspects of video-language understanding. EgoSchema \parencite nips2023egoschema is a video benchmark designed to evaluate the long-form video understanding capabilities within the ego-centric scenario. Derived from Ego4D \parencite cvpr2022ego4d, the benchmark comprises over 5,031 multiple choice question-answer pairs spanning more than 250 hours real-world videos with a semi-automatic data pipeline. VSI-Bench \parencite arxiv2024vsibench is designed to evaluate the visual-spatial comprehensive capabilities of large multimodal models. It consists of over 5,000 question-answer pairs across around 290 real indoor-scene videos. TOMATO \parencite iclr2025tomato is a video benchmark comprises 1,484 human-annotated question-answer pairs and 1,417 videos. TOMATO focuses on evaluating the temporal reasoning capabilities of large multimodal models, including action counting, direction prediction, rotation analysis, shape & trend detection, velocity & frequency estimation, and visual cue interpretation. B.3 Agent Benchmark ScreenSpot V2 \parencite wu2024osatlas is an enhanced version of the ScreenSpot \parencite cheng2024seeclick benchmark, which focuses on evaluating the performance of GUI grounding models across multiple platforms, including web, desktop, and mobile interfaces. This updated version addresses several issues identified in the original ScreenSpot dataset, such as incorrect or ambiguous annotations, spelling mistakes, and mislabeled bounding boxes. ScreenSpot Pro \parencite li2025screenspotpro is a benchmark for evaluating GUI grounding in high-resolution, complex UI environments. It contains 1,581 real-world, high-resolution images and expert-annotated tasks from diverse professional domains. Including domain-specific interface conventions that challenge models to understand professional-grade interfaces beyond consumer applications. OSWorld \parencite xie2024osworld is a pioneering scalable, real computer environment designed for multimodal agents, facilitating task setup, execution-based evaluation, and interactive learning across multiple operating systems, including Ubuntu, Windows, and macOS. It serves as a unified platform for evaluating open-ended computer tasks that involve arbitrary applications, addressing the limitations of existing benchmarks that often lack interactive environments or are confined to specific applications or domains. WindowsAgentArena \parencite bonatti2024windowsagentarenaevaluating is a benchmark designed to evaluate multimodal agents in realistic Windows environments. Built on the OSWorld framework, it allows agents to interact with a full range of applications and web tools. The benchmark is scalable and can complete evaluations in under 20 minutes on Azure. It offers insights into agent performance, highlighting the potential for future research in agent development and task automation.