## Bar Charts: Model Performance Across Benchmarks ### Overview The image presents a series of bar charts comparing the performance of different language models (Nanbeige4-3B, Qwen3-4B-2507, Qwen3-8B, Qwen3-14B, Qwen3-32B, Qwen3-30B-A3B-2507) across various benchmarks. The benchmarks are grouped into four categories: Mathematical Reasoning, Scientific Reasoning, Tool Use & Coding, and Human Preference Alignment. Each category contains two specific benchmarks. The y-axis represents the score achieved on each benchmark, but the scale is not explicitly labeled, implying it's a percentage or a standardized score. ### Components/Axes * **X-axis:** Categorical axis representing the different language models: Nanbeige4-3B, Qwen3-4B-2507, Qwen3-8B, Qwen3-14B, Qwen3-32B, Qwen3-30B-A3B-2507. * **Y-axis:** Numerical axis representing the score achieved by each model on the benchmark. The scale is not explicitly labeled. * **Bar Color:** Teal bars represent the "Nanbeige4-3B" model, while light gray bars represent the "Qwen" models. * **Chart Titles:** Each chart has a title indicating the benchmark category and specific benchmark name. ### Detailed Analysis **1. Mathematical Reasoning** * **AIME 2024:** * Nanbeige4-3B: 90.4 * Qwen3-4B-2507: 83.3 * Qwen3-8B: 76.0 * Qwen3-14B: 79.3 * Qwen3-32B: 81.4 * Qwen3-30B-A3B-2507: 89.2 * Trend: Nanbeige4-3B has the highest score, followed by Qwen3-30B-A3B-2507. Qwen3-8B has the lowest score. * **AIME 2025:** * Nanbeige4-3B: 85.6 * Qwen3-4B-2507: 81.3 * Qwen3-8B: 67.3 * Qwen3-14B: 70.4 * Qwen3-32B: 72.9 * Qwen3-30B-A3B-2507: 85.0 * Trend: Nanbeige4-3B has the highest score, closely followed by Qwen3-30B-A3B-2507. Qwen3-8B has the lowest score. **2. Scientific Reasoning** * **GPQA-Diamond:** * Nanbeige4-3B: 82.2 * Qwen3-4B-2507: 67.2 * Qwen3-8B: 62.0 * Qwen3-14B: 64.0 * Qwen3-32B: 68.7 * Qwen3-30B-A3B-2507: 73.4 * Trend: Nanbeige4-3B significantly outperforms the other models. Qwen3-8B has the lowest score. * **SuperGPQA:** * Nanbeige4-3B: 53.2 * Qwen3-4B-2507: 46.7 * Qwen3-8B: 39.1 * Qwen3-14B: 46.8 * Qwen3-32B: 54.1 * Qwen3-30B-A3B-2507: 56.8 * Trend: Qwen3-30B-A3B-2507 has the highest score, closely followed by Qwen3-32B. Qwen3-8B has the lowest score. **3. Tool Use & Coding** * **BFCL-v4:** * Nanbeige4-3B: 53.8 * Qwen3-4B-2507: 44.9 * Qwen3-8B: 42.2 * Qwen3-14B: 45.4 * Qwen3-32B: 47.9 * Qwen3-30B-A3B-2507: 48.6 * Trend: Nanbeige4-3B has the highest score. Qwen3-8B has the lowest score. * **Fullstack Bench:** * Nanbeige4-3B: 48.0 * Qwen3-4B-2507: 47.1 * Qwen3-8B: 51.5 * Qwen3-14B: 55.7 * Qwen3-32B: 58.2 * Qwen3-30B-A3B-2507: 54.4 * Trend: Qwen3-32B has the highest score. Nanbeige4-3B has one of the lowest scores. **4. Human Preference Alignment** * **ArenaHard-V2:** * Nanbeige4-3B: 60.0 * Qwen3-4B-2507: 40.5 * Qwen3-8B: 26.4 * Qwen3-14B: 39.9 * Qwen3-32B: 48.4 * Qwen3-30B-A3B-2507: 60.0 * Trend: Nanbeige4-3B and Qwen3-30B-A3B-2507 have the highest scores. Qwen3-8B has the lowest score. * **Multi-Challenge:** * Nanbeige4-3B: 41.2 * Qwen3-4B-2507: 41.8 * Qwen3-8B: 35.8 * Qwen3-14B: 36.4 * Qwen3-32B: 39.2 * Qwen3-30B-A3B-2507: 49.4 * Trend: Qwen3-30B-A3B-2507 has the highest score. Qwen3-8B has the lowest score. ### Key Observations * Nanbeige4-3B consistently performs well across most benchmarks, often achieving the highest or near-highest scores. * Qwen3-8B tends to have the lowest scores across many benchmarks. * Qwen3-30B-A3B-2507 shows strong performance, often competing with Nanbeige4-3B. * The performance of the Qwen models varies depending on the specific benchmark. ### Interpretation The data suggests that Nanbeige4-3B is a strong general-purpose language model, demonstrating high performance across a variety of tasks. Qwen3-30B-A3B-2507 also shows promise. The varying performance of the other Qwen models highlights the importance of model selection based on the specific task at hand. The consistent underperformance of Qwen3-8B may indicate limitations in its architecture or training data. The benchmarks cover a range of cognitive abilities, including mathematical reasoning, scientific reasoning, tool use, coding, and human preference alignment, providing a comprehensive evaluation of the models' capabilities.

\n ## Bar Chart: Model Performance on Various Benchmarks ### Overview The image presents a bar chart comparing the performance of several language models (NanBeigel-3B, Qwen-3-4B-2507, Qwen3-14B, Qwen3-30B-A3B-2507) across six different benchmarks: AIME 2024, AIME 2025, GPQA-Diamond, SuperGPQA, BFCL-v4, Fullstack Bench, ArenaHard-v2, and Multi-Challenge. The performance metric appears to be a score, likely accuracy or a similar measure. ### Components/Axes * **X-axis:** Represents the different language models: NanBeigel-3B, Qwen-3-4B-2507, Qwen3-14B, Qwen3-30B-A3B-2507. * **Y-axis:** Represents the performance score, ranging from approximately 20 to 90. The scale is not explicitly labeled, but can be inferred from the values displayed on the bars. * **Chart Title:** The chart is divided into sections, each with a title indicating the benchmark being evaluated: Mathematical Reasoning (AIME 2024, AIME 2025), Scientific Reasoning (GPQA-Diamond, SuperGPQA), Tool Use & Coding (BFCL-v4, Fullstack Bench), and Human Preference Alignment (ArenaHard-v2, Multi-Challenge). * **Bars:** Each bar represents the performance of a specific model on a specific benchmark. The color of the bars is consistent within each benchmark section. ### Detailed Analysis **Mathematical Reasoning** * **AIME 2024:** * NanBeigel-3B: Approximately 90.4 * Qwen-3-4B-2507: Approximately 83.3 * Qwen3-14B: Approximately 79.3 * Qwen3-30B-A3B-2507: Approximately 81.4, 76.0 * **AIME 2025:** * NanBeigel-3B: Approximately 89.2 * Qwen-3-4B-2507: Approximately 81.3 * Qwen3-14B: Approximately 70.4 * Qwen3-30B-A3B-2507: Approximately 72.9, 85.0 **Scientific Reasoning** * **GPQA-Diamond:** * NanBeigel-3B: Approximately 82.2 * Qwen-3-4B-2507: Approximately 67.2 * Qwen3-14B: Approximately 62.0 * Qwen3-30B-A3B-2507: Approximately 64.0, 68.7 * **SuperGPQA:** * NanBeigel-3B: Approximately 53.2 * Qwen-3-4B-2507: Approximately 46.7 * Qwen3-14B: Approximately 39.1 * Qwen3-30B-A3B-2507: Approximately 54.1, 50.8, 46.8 **Tool Use & Coding** * **BFCL-v4:** * NanBeigel-3B: Approximately 53.8 * Qwen-3-4B-2507: Approximately 44.9 * Qwen3-14B: Approximately 45.4 * Qwen3-30B-A3B-2507: Approximately 47.9, 48.6, 42.2 * **Fullstack Bench:** * NanBeigel-3B: Approximately 58.2 * Qwen-3-4B-2507: Approximately 48.0 * Qwen3-14B: Approximately 47.1 * Qwen3-30B-A3B-2507: Approximately 51.5, 54.4 **Human Preference Alignment** * **ArenaHard-v2:** * NanBeigel-3B: Approximately 60.0 * Qwen-3-4B-2507: Approximately 40.5 * Qwen3-14B: Approximately 39.9 * Qwen3-30B-A3B-2507: Approximately 26.4 * **Multi-Challenge:** * NanBeigel-3B: Approximately 49.4 * Qwen-3-4B-2507: Approximately 41.2 * Qwen3-14B: Approximately 35.8 * Qwen3-30B-A3B-2507: Approximately 36.4, 39.2 ### Key Observations * **NanBeigel-3B consistently outperforms other models** across most benchmarks, particularly in Mathematical and Scientific Reasoning. * **Qwen3-30B-A3B-2507 often shows the second-best performance**, frequently surpassing Qwen-3-4B-2507 and Qwen3-14B. * **Qwen3-14B generally performs worse than Qwen-3-4B-2507**, except in some instances within the Tool Use & Coding category. * **Performance varies significantly across benchmarks.** Models that excel in one area may not perform as well in others. * The presence of multiple bars for Qwen3-30B-A3B-2507 in some benchmarks suggests multiple runs or variations of the model were evaluated. ### Interpretation The data suggests that NanBeigel-3B is the strongest performing model among those tested, demonstrating superior capabilities in mathematical, scientific, and general reasoning tasks. Qwen3-30B-A3B-2507 represents a significant step up in performance compared to the smaller Qwen models, indicating that increasing model size generally leads to improved results. The varying performance across benchmarks highlights the importance of evaluating models on a diverse set of tasks to gain a comprehensive understanding of their strengths and weaknesses. The multiple bars for Qwen3-30B-A3B-2507 could indicate the impact of different training configurations or data subsets on the model's performance. The relatively lower scores on the Human Preference Alignment benchmarks suggest that these tasks may be more challenging for the models, or that the evaluation metrics do not fully capture human preferences. Overall, the chart provides valuable insights into the relative performance of these language models and can inform future research and development efforts.

## Bar Chart Composite: AI Model Benchmark Performance ### Overview The image displays a composite of eight bar charts, organized into four thematic categories, comparing the performance of six different AI models across various standardized benchmarks. The models compared are: **Nanbeige4-3B** (highlighted in teal), **Qwen3-4B-2507**, **Qwen3-8B**, **Qwen3-14B**, **Qwen3-32B**, and **Qwen3-30B-A3B-2507** (all in shades of gray). The charts are grouped under the headings: Mathematical Reasoning, Scientific Reasoning, Tool Use & Coding, and Human Preference Alignment. ### Components/Axes * **Chart Structure:** Eight individual bar charts arranged in a 2x4 grid. * **Categories (Top Headers):** * Top Left: **Mathematical Reasoning** * Top Right: **Scientific Reasoning** * Bottom Left: **Tool Use & Coding** * Bottom Right: **Human Preference Alignment** * **Sub-Charts (Benchmark Titles):** * Under Mathematical Reasoning: **AIME 2024**, **AIME 2025** * Under Scientific Reasoning: **GPQA-Diamond**, **SuperGPQA** * Under Tool Use & Coding: **BFCL-v4**, **Fullstack Bench** * Under Human Preference Alignment: **ArenaHard-V2**, **Multi-Challenge** * **X-Axis (All Charts):** Lists the six model names. The labels are rotated approximately 45 degrees for readability. * **Y-Axis (All Charts):** Represents the benchmark score. The scale is not explicitly numbered with ticks, but the numerical value is printed directly above each bar. * **Legend/Color Key:** * **Teal Bar:** Nanbeige4-3B * **Gray Bars (from left to right):** Qwen3-4B-2507, Qwen3-8B, Qwen3-14B, Qwen3-32B, Qwen3-30B-A3B-2507. ### Detailed Analysis #### **Mathematical Reasoning** 1. **AIME 2024:** * Nanbeige4-3B: **90.4** (Highest) * Qwen3-4B-2507: **83.3** * Qwen3-8B: **76.0** * Qwen3-14B: **79.3** * Qwen3-32B: **81.4** * Qwen3-30B-A3B-2507: **89.2** (Second highest) * *Trend:* Nanbeige4-3B leads, followed closely by the largest Qwen model (30B-A3B). Performance dips for the mid-sized Qwen models (8B, 14B, 32B). 2. **AIME 2025:** * Nanbeige4-3B: **85.6** (Highest) * Qwen3-4B-2507: **81.3** * Qwen3-8B: **67.3** * Qwen3-14B: **70.4** * Qwen3-32B: **72.9** * Qwen3-30B-A3B-2507: **85.0** (Very close second) * *Trend:* Similar pattern to AIME 2024. Nanbeige4-3B and Qwen3-30B-A3B-2507 are nearly tied at the top, with a significant drop for the 8B, 14B, and 32B models. #### **Scientific Reasoning** 1. **GPQA-Diamond:** * Nanbeige4-3B: **82.2** (Highest) * Qwen3-4B-2507: **67.2** * Qwen3-8B: **62.0** * Qwen3-14B: **64.0** * Qwen3-32B: **68.7** * Qwen3-30B-A3B-2507: **73.4** * *Trend:* Clear lead for Nanbeige4-3B. Performance generally increases with model size within the Qwen series, but all are below Nanbeige4-3B. 2. **SuperGPQA:** * Nanbeige4-3B: **53.2** (Highest) * Qwen3-4B-2507: **46.7** * Qwen3-8B: **39.1** * Qwen3-14B: **46.8** * Qwen3-32B: **54.1** (Slightly higher than Nanbeige4-3B) * Qwen3-30B-A3B-2507: **56.8** (Highest) * *Trend:* This is the only benchmark where a Qwen model (Qwen3-30B-A3B-2507) clearly outperforms Nanbeige4-3B. Qwen3-32B also scores slightly higher. The smallest Qwen model (8B) performs notably worse. #### **Tool Use & Coding** 1. **BFCL-v4:** * Nanbeige4-3B: **53.8** (Highest) * Qwen3-4B-2507: **44.9** * Qwen3-8B: **42.2** * Qwen3-14B: **45.4** * Qwen3-32B: **47.9** * Qwen3-30B-A3B-2507: **48.6** * *Trend:* Nanbeige4-3B holds a clear lead. Performance among Qwen models improves with scale but remains below Nanbeige4-3B. 2. **Fullstack Bench:** * Nanbeige4-3B: **48.0** * Qwen3-4B-2507: **47.1** * Qwen3-8B: **51.5** * Qwen3-14B: **55.7** * Qwen3-32B: **58.2** (Highest) * Qwen3-30B-A3B-2507: **54.4** * *Trend:* This benchmark shows a different pattern. Nanbeige4-3B is not the leader. Performance generally increases with Qwen model size, peaking at Qwen3-32B. #### **Human Preference Alignment** 1. **ArenaHard-V2:** * Nanbeige4-3B: **60.0** (Tied for Highest) * Qwen3-4B-2507: **40.5** * Qwen3-8B: **26.4** (Lowest across all charts) * Qwen3-14B: **39.9** * Qwen3-32B: **48.4** * Qwen3-30B-A3B-2507: **60.0** (Tied for Highest) * *Trend:* Nanbeige4-3B and Qwen3-30B-A3B-2507 are tied at the top. There is a very sharp drop for the Qwen3-8B model. 2. **Multi-Challenge:** * Nanbeige4-3B: **41.2** * Qwen3-4B-2507: **41.8** (Slightly higher than Nanbeige4-3B) * Qwen3-8B: **35.8** * Qwen3-14B: **36.4** * Qwen3-32B: **39.2** * Qwen3-30B-A3B-2507: **49.4** (Highest) * *Trend:* Qwen3-30B-A3B-2507 is the clear leader. Nanbeige4-3B is competitive but slightly outperformed by the smaller Qwen3-4B-2507. ### Key Observations 1. **Nanbeige4-3B Dominance:** The Nanbeige4-3B model (teal) achieves the highest score in 6 out of the 8 benchmarks presented (AIME 2024, AIME 2025, GPQA-Diamond, SuperGPQA, BFCL-v4, ArenaHard-V2). 2. **Strongest Competitor:** The **Qwen3-30B-A3B-2507** model is the most consistent high performer among the Qwen series, often coming in a close second or even surpassing Nanbeige4-3B (as in SuperGPQA, Fullstack Bench, and Multi-Challenge). 3. **Performance vs. Scale (Qwen Series):** Within the Qwen models, performance generally improves with increasing parameter size (from 4B to 32B), but this is not perfectly linear. The 30B-A3B-2507 variant often outperforms the standard 32B model. 4. **Notable Low Point:** The **Qwen3-8B** model shows a significant performance dip, particularly in the ArenaHard-V2 benchmark where it scores only 26.4, the lowest value in the entire composite. 5. **Benchmark Variability:** No single model dominates every category. The relative performance shifts between mathematical, scientific, coding, and alignment-focused tasks. ### Interpretation This composite chart provides a comparative snapshot of AI model capabilities across a diverse evaluation suite. The data suggests that **Nanbeige4-3B** is a highly capable and well-rounded model, excelling particularly in reasoning-heavy tasks (Math, Science, GPQA). Its strength in the ArenaHard-V2 benchmark also indicates good alignment with human preferences. The **Qwen3 series** demonstrates a clear scaling law, with larger models generally performing better. The **Qwen3-30B-A3B-2507** variant appears to be a particularly efficient or well-tuned architecture, as it frequently matches or beats the larger standard 32B model and competes directly with Nanbeige4-3B. Its victory in the **Multi-Challenge** benchmark suggests superior handling of diverse, complex tasks. The poor showing of **Qwen3-8B** on ArenaHard-V2 is an outlier that may indicate a specific weakness in that model's alignment or a mismatch with that particular benchmark's evaluation criteria. Overall, the charts illustrate that model performance is highly task-dependent, and choosing the "best" model requires considering the specific application domain (e.g., math vs. coding vs. general assistant tasks). The visualization effectively communicates these nuanced comparisons through clear, direct score labeling and consistent color coding.

## Bar Chart: AI Model Performance Across Benchmarks ### Overview The image presents a comparative analysis of six AI models across multiple reasoning and capability benchmarks. The chart is divided into four main sections: Mathematical Reasoning, Scientific Reasoning, Tool Use & Coding, and Human Preference Alignment. Each section contains two subsections comparing performance across different years (2024 vs. 2025) or benchmark types (e.g., AIME vs. GPQA). ### Components/Axes - **X-axis**: AI models (Nanbeige4-3B, Gwen3-4B-2507, Gwen3-3B-8B, Gwen3-3B-14B, Gwen3-3B-32B, Gwen3-30B-A3B-2507) - **Y-axis**: Performance scores (percentage values) - **Legend**: Color-coded model identifiers (teal, gray, white) - **Sections**: 1. Mathematical Reasoning (AIME 2024/2025) 2. Scientific Reasoning (GPQA-Diamond/SuperGPQA) 3. Tool Use & Coding (BFCL-v4/Fullstack Bench) 4. Human Preference Alignment (ArenaHard-V2/Multi-Challenge) ### Detailed Analysis #### Mathematical Reasoning - **AIME 2024**: - Nanbeige4-3B: 90.4% (teal) - Gwen3-4B-2507: 83.3% (gray) - Gwen3-3B-8B: 76.0% (white) - Gwen3-3B-14B: 79.3% (gray) - Gwen3-3B-32B: 81.4% (gray) - Gwen3-30B-A3B-2507: 89.2% (gray) - **AIME 2025**: - Nanbeige4-3B: 85.6% (teal) - Gwen3-4B-2507: 81.3% (gray) - Gwen3-3B-8B: 67.3% (white) - Gwen3-3B-14B: 70.4% (gray) - Gwen3-3B-32B: 72.9% (gray) - Gwen3-30B-A3B-2507: 85.0% (gray) #### Scientific Reasoning - **GPQA-Diamond**: - Nanbeige4-3B: 82.2% (teal) - Gwen3-4B-2507: 67.2% (gray) - Gwen3-3B-8B: 62.0% (white) - Gwen3-3B-14B: 64.0% (gray) - Gwen3-3B-32B: 68.7% (gray) - Gwen3-30B-A3B-2507: 73.4% (gray) - **SuperGPQA**: - Nanbeige4-3B: 53.2% (teal) - Gwen3-4B-2507: 46.7% (gray) - Gwen3-3B-8B: 39.1% (white) - Gwen3-3B-14B: 46.8% (gray) - Gwen3-3B-32B: 54.1% (gray) - Gwen3-30B-A3B-2507: 56.8% (gray) #### Tool Use & Coding - **BFCL-v4**: - Nanbeige4-3B: 53.8% (teal) - Gwen3-4B-2507: 44.9% (gray) - Gwen3-3B-8B: 42.2% (white) - Gwen3-3B-14B: 45.4% (gray) - Gwen3-3B-32B: 47.9% (gray) - Gwen3-30B-A3B-2507: 48.6% (gray) - **Fullstack Bench**: - Nanbeige4-3B: 48.0% (teal) - Gwen3-4B-2507: 47.1% (gray) - Gwen3-3B-8B: 51.5% (white) - Gwen3-3B-14B: 55.7% (gray) - Gwen3-3B-32B: 58.2% (gray) - Gwen3-30B-A3B-2507: 54.4% (gray) #### Human Preference Alignment - **ArenaHard-V2**: - Nanbeige4-3B: 60.0% (teal) - Gwen3-4B-2507: 40.5% (gray) - Gwen3-3B-8B: 26.4% (white) - Gwen3-3B-14B: 39.9% (gray) - Gwen3-3B-32B: 48.4% (gray) - Gwen3-30B-A3B-2507: 60.0% (gray) - **Multi-Challenge**: - Nanbeige4-3B: 41.2% (teal) - Gwen3-4B-2507: 41.8% (gray) - Gwen3-3B-8B: 35.8% (white) - Gwen3-3B-14B: 36.4% (gray) - Gwen3-3B-32B: 39.2% (gray) - Gwen3-30B-A3B-2507: 49.4% (gray) ### Key Observations 1. **2025 Models Outperform 2024 in Most Technical Tasks**: - Gwen3-30B-A3B-2507 consistently achieves the highest scores across all benchmarks except ArenaHard-V2. - AIME 2025 shows a 4.2% average improvement over AIME 2024 for Gwen3 models. 2. **Human Preference Alignment Anomaly**: - Gwen3-30B-A3B-2507 matches Nanbeige4-3B in ArenaHard-V2 (60.0%) but underperforms in Multi-Challenge (49.4% vs. 41.2%). - Gwen3-3B-8B shows the largest drop in Human Preference Alignment (26.4% vs. 35.8% in Multi-Challenge). 3. **Tool Use & Coding Trends**: - Fullstack Bench scores are 10-15% higher than BFCL-v4 across all models. - Gwen3-3B-32B demonstrates the steepest improvement in Fullstack Bench (58.2% vs. 47.9% in BFCL-v4). ### Interpretation The data suggests that newer Gwen3 models (2025) exhibit significant improvements in technical reasoning and coding capabilities compared to their 2024 counterparts. However, the Human Preference Alignment section reveals a critical divergence: while Gwen3-30B-A3B-2507 matches Nanbeige4-3B in ArenaHard-V2, its Multi-Challenge score (49.4%) remains below Nanbeige4-3B's 41.2%, indicating potential misalignment with human expectations in complex scenarios. This could imply that while technical capabilities have advanced, value alignment and nuanced decision-making remain challenges for larger models. The consistent performance of Gwen3-3B-32B in Fullstack Bench (58.2%) suggests specialized training in coding tasks yields disproportionate gains compared to general reasoning improvements.