## Scatter Plot: AIME 2025 Performance vs. Total Parameters ### Overview The image is a scatter plot comparing the AIME 2025 Pass@1 Score (y-axis) against the Total Parameters (x-axis) for various language models. The plot uses different colored markers to distinguish between models with "Open-Weights Only" and "Open-Weights & Open-Data". A horizontal dashed line indicates a performance threshold. ### Components/Axes * **Title:** AIME 2025 Performance vs. Total Parameters * **X-axis:** Total Parameters, with a logarithmic scale. Markers at 4B, 10B, 32B, and 100B. * **Y-axis:** AIME 2025 Pass@1 Score, with a linear scale. Markers at 50, 55, 60, 65, 70, 75, 80, 85, and 90. * **Legend (bottom-right):** * Gray circle: Open-Weights Only * Tan circle: Open-Weights & Open-Data * Orange star: Our Model * **Horizontal Dashed Line:** Located at approximately 83 on the y-axis. ### Detailed Analysis The data points are scattered across the plot, showing the relationship between model size (Total Parameters) and performance (AIME 2025 Pass@1 Score). * **DASD-4B-Thinking (Ours):** Marked with an orange star, located at approximately (4B, 84). This is the highest performing model. * **POLARIS-4B:** Gray circle, located at approximately (4B, 80). * **Qwen3-4B-Thinking:** Tan circle, located at approximately (5B, 80). * **Mistral3-8B:** Gray circle, located at approximately (8B, 78). * **Nvidia-OpenReasoning-7B:** Tan circle, located at approximately (7B, 77). * **DeepSeek-R1-Qwen3-8B:** Gray circle, located at approximately (8B, 74). * **Mistral3-3B:** Gray circle, located at approximately (4B, 72). * **Qwen3-14B:** Gray circle, located at approximately (14B, 70). * **AM-thinking-v1:** Tan circle, located at approximately (30B, 73). * **Qwen3-32B:** Gray circle, located at approximately (30B, 72). * **Nvidia-Nemotron-Ultra-253B:** Tan circle, located at approximately (80B, 73). This marker is larger than the others. * **GLM-Z1-32B:** Gray circle, located at approximately (30B, 60). * **GLM-Z1-9B:** Gray circle, located at approximately (9B, 57). * **OpenThoughts3-7B:** Tan circle, located at approximately (7B, 53). ### Key Observations * The "Our Model" (DASD-4B-Thinking) significantly outperforms other models in terms of AIME 2025 Pass@1 Score, despite having a relatively small number of parameters (4B). * There isn't a clear linear correlation between the number of parameters and the AIME 2025 Pass@1 Score. Some models with fewer parameters achieve higher scores than models with more parameters. * The models using "Open-Weights & Open-Data" (tan circles) are scattered across the plot, with some performing better than others. * The size of the data point for "Nvidia-Nemotron-Ultra-253B" is larger than the other data points. ### Interpretation The scatter plot suggests that model performance on the AIME 2025 benchmark is not solely determined by the number of parameters. Factors such as model architecture, training data, and training methods likely play a significant role. The "Our Model" data point indicates that efficient design or specialized training can lead to superior performance even with fewer parameters. The horizontal line may represent a target performance threshold, which only "Our Model" exceeds. The size of the "Nvidia-Nemotron-Ultra-253B" data point may indicate the relative size of the model, or some other factor.

## Scatter Plot: AIME 2025 Performance vs. Total Parameters ### Overview This scatter plot visualizes the relationship between the AIME 2025 Pass@1 Score and the Total Parameters of various language models. The plot displays model performance on the y-axis (AIME 2025 Pass@1 Score) against model size on the x-axis (Total Parameters). Models are color-coded based on their category (Open-Weights Only, Open-Weights & Open-Data, and Our Model). ### Components/Axes * **Title:** AIME 2025 Performance vs. Total Parameters * **X-axis:** Total Parameters (Scale: 4B, 10B, 32B, 100B) * **Y-axis:** AIME 2025 Pass@1 Score (Scale: 50, 60, 70, 80, 90) * **Legend:** * Gray Circle: Open-Weights Only * Yellow Circle: Open-Weights & Open-Data * Red Star: Our Model * **Data Points:** Represent individual language models, positioned according to their AIME 2025 Pass@1 Score and Total Parameters. ### Detailed Analysis The data points are scattered across the plot, showing a general trend of increasing performance with increasing parameters, but with significant variation. Here's a breakdown of the data points, reading from left to right (increasing parameter count): * **4B:** * POLARIS-4B: Approximately (4B, 81). Gray circle. * Mistral3-3B: Approximately (4B, 71). Yellow circle. * **10B:** * Nvidia-OpenReasoning-7B: Approximately (10B, 76). Yellow circle. * DeepSeek-R1-Qwen3-8B: Approximately (10B, 74). Yellow circle. * Mistral3-8B: Approximately (10B, 80). Yellow circle. * Qwen3-4B-Thinking: Approximately (10B, 80). Gray circle. * GLM-ZI-9B: Approximately (10B, 56). Yellow circle. * **32B:** * Qwen3-14B: Approximately (32B, 70). Yellow circle. * Qwen3-32B: Approximately (32B, 73). Yellow circle. * GLM-ZI-32B: Approximately (32B, 60). Yellow circle. * **100B:** * AM-thinking-v1: Approximately (100B, 74). Yellow circle. * Nvidia-Nemotron-Ultra-253B: Approximately (100B, 72). Yellow circle. * **Our Model:** * DASD-4B-Thinking (Ours): Approximately (4B, 84). Red star. * OpenThoughts3-7B: Approximately (10B, 52). Yellow circle. The trend for the yellow circles (Open-Weights & Open-Data) is generally upward, but with considerable spread. The gray circles (Open-Weights Only) are clustered around the 80 score mark. Our model (red star) shows a high score at a relatively small parameter size. ### Key Observations * **Outlier:** DASD-4B-Thinking (Our Model) significantly outperforms other models with a similar parameter count (4B). * **Parameter Scaling:** There's a positive correlation between parameters and performance, but the relationship isn't strictly linear. Some models with fewer parameters perform comparably to those with more. * **Model Category Variation:** The Open-Weights & Open-Data models (yellow) exhibit the widest range of performance. * **Clustering:** Models in the 4B and 10B range show a tighter clustering of performance scores. ### Interpretation The data suggests that model size (Total Parameters) is a significant factor in achieving high AIME 2025 Pass@1 Scores, but it is not the sole determinant. The category of the model (Open-Weights Only vs. Open-Weights & Open-Data) also plays a role, with the latter exhibiting more variability. The outlier performance of "DASD-4B-Thinking (Ours)" indicates that architectural innovations or training methodologies can lead to substantial gains in performance even with limited parameters. The spread of data points highlights the importance of factors beyond model size, such as training data quality, model architecture, and optimization techniques. The plot provides a comparative snapshot of the current landscape of language models, demonstrating the trade-offs between model size, performance, and accessibility (as indicated by the Open-Weights/Open-Data categorization). The data suggests that focusing solely on increasing parameters may not be the most effective strategy for improving performance; rather, a holistic approach that considers all aspects of model development is crucial.

## Scatter Plot: AIME 2025 Performance vs. Total Parameters ### Overview This is a scatter plot comparing the performance of various AI models on the AIME 2025 benchmark against their total parameter count. The chart highlights a specific model, "DASD-4B-Thinking (Ours)," and compares it against other models categorized by their weight and data openness. ### Components/Axes * **Chart Title:** "AIME 2025 Performance vs. Total Parameters" * **Y-Axis:** "AIME 2025 Pass@1 Score". The scale runs from 50 to 90, with major tick marks every 5 units (50, 55, 60, 65, 70, 75, 80, 85, 90). * **X-Axis:** "Total Parameters". The scale is logarithmic, with labeled tick marks at 4B, 10B, 32B, and 100B (where B likely denotes billion). * **Legend:** Located in the bottom-right quadrant. It defines three model categories: * **Gray Circle:** "Open-Weights Only" * **Tan/Gold Circle:** "Open-Weights & Open-Data" * **Orange Star:** "Our Model" * **Reference Line:** A horizontal, dashed orange line extends from the "Our Model" data point across the chart at a score of approximately 83.5. ### Detailed Analysis The plot contains 13 distinct data points. Below is a list of each model, its approximate parameter count (x-axis), its approximate AIME score (y-axis), and its category based on the legend. **Models with ~4B Parameters:** 1. **DASD-4B-Thinking (Ours):** ~4B parameters, Score: ~83.5. (Orange Star - "Our Model"). This is the highest-performing model on the chart. 2. **POLARIS-4B:** ~4B parameters, Score: ~79. (Tan Circle - "Open-Weights & Open-Data"). 3. **Mistral3-3B:** ~3B parameters, Score: ~72. (Gray Circle - "Open-Weights Only"). **Models with ~7B-8B Parameters:** 4. **Qwen3-4B-Thinking:** ~7B parameters, Score: ~79. (Gray Circle - "Open-Weights Only"). 5. **Nvidia-OpenReasoning-7B:** ~7B parameters, Score: ~77. (Gray Circle - "Open-Weights Only"). 6. **Mistral3-8B:** ~8B parameters, Score: ~78. (Gray Circle - "Open-Weights Only"). 7. **DeepSeek-R1-Qwen3-8B:** ~8B parameters, Score: ~76. (Gray Circle - "Open-Weights Only"). 8. **OpenThoughts3-7B:** ~7B parameters, Score: ~53. (Tan Circle - "Open-Weights & Open-Data"). This is a notable low-performer for its size. 9. **GLM-Z1-9B:** ~9B parameters, Score: ~56. (Gray Circle - "Open-Weights Only"). **Models with ~14B-32B Parameters:** 10. **Qwen3-14B:** ~14B parameters, Score: ~70. (Gray Circle - "Open-Weights Only"). 11. **GLM-Z1-32B:** ~32B parameters, Score: ~63. (Gray Circle - "Open-Weights Only"). 12. **Qwen3-32B:** ~32B parameters, Score: ~72. (Gray Circle - "Open-Weights Only"). 13. **AM-thinking-v1:** ~32B parameters, Score: ~74. (Tan Circle - "Open-Weights & Open-Data"). **Model with ~253B Parameters:** 14. **Nvidia-Nemotron-Ultra-253B:** ~253B parameters, Score: ~72. (Tan Circle - "Open-Weights & Open-Data"). ### Key Observations 1. **Performance Outlier:** The "Our Model" (DASD-4B-Thinking) significantly outperforms all other models, achieving the highest score (~83.5) with a relatively small parameter count (~4B). 2. **Efficiency Trend:** There is no clear positive correlation between parameter count and performance. Several smaller models (e.g., POLARIS-4B, Qwen3-4B-Thinking) outperform much larger models (e.g., Nvidia-Nemotron-Ultra-253B, Qwen3-32B). 3. **Low-Performing Cluster:** Two models, OpenThoughts3-7B and GLM-Z1-9B, form a distinct low-scoring cluster with scores in the 50s, despite having parameter counts similar to mid-performing 7B-8B models. 4. **Category Distribution:** "Open-Weights Only" models (gray) are the most numerous and show the widest performance variance (scores from ~56 to ~79). "Open-Weights & Open-Data" models (tan) are fewer but include both high performers (POLARIS-4B) and the largest model on the chart. ### Interpretation The data suggests that model architecture, training methodology, or data quality (factors not directly shown) are more critical to achieving high AIME 2025 scores than sheer model size. The standout performance of "DASD-4B-Thinking" implies a significant efficiency breakthrough, achieving state-of-the-art results with a compact model. The chart effectively argues that bigger is not always better in this benchmark context. The presence of low-scoring models in the 7B-9B range indicates that this parameter size is a competitive but volatile space where implementation details lead to drastically different outcomes. The horizontal reference line from the "Our Model" point serves as a visual benchmark, emphasizing the performance gap it has established over both smaller and substantially larger competitors.

## Scatter Plot: AIME 2025 Performance vs. Total Parameters ### Overview The image is a scatter plot comparing the performance of various AI models on the AIME 2025 benchmark against their total parameter counts. The y-axis represents the "AIME 2025 Pass@1 Score" (ranging from 50 to 90), while the x-axis shows "Total Parameters" (from 4B to 100B). Data points are color-coded by model category: gray for "Open-Weights Only," beige for "Open-Weights & Open-Data," and orange for "Our Model." A red dashed line at ~85 marks a performance threshold. ### Components/Axes - **Title**: "AIME 2025 Performance vs. Total Parameters" (top center). - **X-axis**: "Total Parameters" (logarithmic scale: 4B, 10B, 32B, 100B). - **Y-axis**: "AIME 2025 Pass@1 Score" (linear scale: 50–90). - **Legend**: Bottom-right corner, with three categories: - Gray: Open-Weights Only - Beige: Open-Weights & Open-Data - Orange Star: Our Model - **Data Points**: Labeled with model names and parameter counts (e.g., "DASD-4B-Thinking (Ours)"). ### Detailed Analysis - **Model Categories**: - **Open-Weights Only (Gray)**: - POLARIS-4B (4B parameters, ~80 score) - Mistral3-3B (4B, ~72) - Nvidia-OpenReasoning-7B (7B, ~75) - DeepSeek-R1-Qwen3-8B (8B, ~75) - Qwen3-14B (14B, ~70) - GLM-Z1-32B (32B, ~65) - OpenThoughts3-7B (7B, ~55) - GLM-Z1-9B (9B, ~55) - **Open-Weights & Open-Data (Beige)**: - Qwen3-4B-Thinking (4B, ~80) - Mistral3-8B (8B, ~78) - Qwen3-32B (32B, ~74) - Nvidia-Nemotron-Ultra-253B (253B, ~73) - AM-thinking-v1 (32B, ~74) - **Our Model (Orange Star)**: - DASD-4B-Thinking (4B, ~85). - **Trends**: - **Performance vs. Parameters**: Larger models generally achieve higher scores, but exceptions exist (e.g., Qwen3-14B at 14B parameters scores lower than Mistral3-8B at 8B). - **Category Performance**: "Open-Weights & Open-Data" models (beige) dominate the upper-right quadrant, while "Open-Weights Only" (gray) are more spread out. "Our Model" (DASD-4B-Thinking) achieves the highest score despite having the fewest parameters. ### Key Observations 1. **Outlier**: DASD-4B-Thinking (orange star) outperforms all other models, including larger ones, suggesting efficiency or architectural advantages. 2. **Threshold**: The red dashed line at ~85 indicates a performance benchmark; only DASD-4B-Thinking and Qwen3-4B-Thinking exceed it. 3. **Parameter Efficiency**: Some smaller models (e.g., POLARIS-4B, Qwen3-4B-Thinking) achieve scores comparable to larger models (e.g., Qwen3-32B, Nvidia-Nemotron-Ultra-253B). ### Interpretation The data highlights a trade-off between model size and performance. While larger models (e.g., 253B parameters) generally perform better, the "Our Model" (DASD-4B-Thinking) demonstrates that optimized architectures or training strategies can achieve superior results with fewer parameters. The dominance of "Open-Weights & Open-Data" models in the high-performance quadrant suggests that open-data integration may enhance performance. However, the variability in scores across similar parameter ranges (e.g., 8B vs. 14B) indicates that factors beyond parameter count—such as training data quality, model architecture, or fine-tuning—play critical roles in benchmark performance.