Best model for math
Models ranked by their math score, 0 to 100, as on the category page. Only ranked models with enough scored items are included, and models are grouped by total parameters.
Read the memory figure as an estimate, not a guarantee. The memory a model needs is estimated from its file: the weights plus about 20% for the context of a single conversation. It is not a measurement, and a long conversation or several at once needs more. A model with no file size on record is not counted as fitting any amount of memory. Speeds, and the GPU memory shown on the speed page, were measured on this benchmark's server, not on yours. That server kept 1 to 8 answers in flight at a time and loaded a context window of 16,384 tokens for most models; 42 of the models ranked ran at a shorter context. Speeds were measured on one laptop and are comparable only between models on the same backend. See the settings and the memory method.
All sizes
Flint 40B A2B leads here with 100.0 on a scale of 0 to 100; its Overall is 89.0 (95% interval 85.1–92.6).
| Place | Model | Math (0 to 100) | Overall | Rank range | Decode (tok/s) | Memory needed (GB, est.) | Size class |
|---|---|---|---|---|---|---|---|
| 1 | Flint 40B A2B | 100.025 of 25 items scored | 89.085.1–92.6 | 14–90 | 25.9ROCm 7.1 | 27.6 | > 35B |
| 1 | Gemma 4 12B IT | 100.025 of 25 items scored | 98.697.0–99.8 | 1–12 | 8.1ROCm 7.1 | 8.9 | 9–16B |
| 1 | gemma-4-26B-A4B-it | 100.025 of 25 items scored | 97.895.8–99.3 | 1–22 | 15.7ROCm 7.1 | 21.6 | 16–35B |
| 1 | gemma-4-31B-it | 100.025 of 25 items scored | 99.398.3–100.0 | 1–5 | 3.8ROCm 7.1 | 22.4 | 16–35B |
| 1 | Muse Glimmer 30B | 100.025 of 25 items scored | 97.996.0–99.3 | 1–23 | 3.6ROCm 7.1 | 21.8 | 16–35B |
Tiny (under 3B)
Deepscaler 1.5B Preview leads its size class with 92.0 on a scale of 0 to 100; its Overall is 52.1 (95% interval 46.6–57.7).
| Place | Model | Math (0 to 100) | Overall | Rank range | Decode (tok/s) | Memory needed (GB, est.) | Size class |
|---|---|---|---|---|---|---|---|
| 1 | Deepscaler 1.5B Previewfine-tune of DeepSeek-R1-Distill-Qwen 1.5Bmath model | 92.025 of 25 items scored | 52.146.6–57.7 | 168–198 | 45.5ROCm 7.1 | 1.3 | Tiny (under 3B) |
| 2 | Deepcoder 1.5B Previewfine-tune of DeepSeek-R1-Distill-Qwen 1.5Bcode model | 88.025 of 25 items scored | 55.850.3–61.0 | 155–192 | 44.9ROCm 7.1 | 1.3 | Tiny (under 3B) |
| 3 | Lfm2.5 Thinking 1.2B | 80.025 of 25 items scored | 74.369.5–79.1 | 84–148 | 205.0ROCm 7.1 | 0.9 | Tiny (under 3B) |
| 3 | Qwen3.5 2B | 80.025 of 25 items scored | 75.169.7–80.3 | 81–144 | 33.3ROCm 7.1 | 2.3 | Tiny (under 3B) |
| 5 | Deepseek R1 1.5B Qwen Distill | 76.025 of 25 items scored | 48.943.5–54.7 | 177–205 | 44.9ROCm 7.1 | 1.3 | Tiny (under 3B) |
3–4B
Phi4 Mini Reasoning 3.8B leads its size class with 96.0 on a scale of 0 to 100; its Overall is 70.9 (95% interval 65.9–76.0).
| Place | Model | Math (0 to 100) | Overall | Rank range | Decode (tok/s) | Memory needed (GB, est.) | Size class |
|---|---|---|---|---|---|---|---|
| 1 | Phi4 Mini Reasoning 3.8B | 96.025 of 25 items scored | 70.965.9–76.0 | 97–157 | 48.8ROCm 7.1 | 3.8 | 3–4B |
| 2 | Ministral 3 3B Instruct 2512 | 88.025 of 25 items scored | 80.976.5–85.2 | 61–124 | 70.9ROCm 7.1 | 3.6 | 3–4B |
| 3 | Spark-X2.6-4B | 84.025 of 25 items scored | 85.080.9–89.1 | 43–106 | 19.7ROCm 7.1 | 3.0 | 3–4B |
| 4 | granite-4.1-3b | 80.025 of 25 items scored | 80.775.9–85.3 | 66–129 | 20.1ROCm 7.1 | 2.5 | 3–4B |
| 5 | Phi-4-mini-instruct | 76.025 of 25 items scored | 81.677.1–86.0 | 62–123 | 17.1ROCm 7.1 | 3.0 | 3–4B |
4–9B
Deepseek R1 8B Llama Distill leads its size class with 96.0 on a scale of 0 to 100; its Overall is 81.7 (95% interval 76.9–86.1).
| Place | Model | Math (0 to 100) | Overall | Rank range | Decode (tok/s) | Memory needed (GB, est.) | Size class |
|---|---|---|---|---|---|---|---|
| 1 | Deepseek R1 8B Llama Distill | 96.025 of 25 items scored | 81.776.9–86.1 | 56–122 | 11.9ROCm 7.1 | 5.9 | 4–9B |
| 1 | DeepSeek-R1-Distill-Qwen-7B | 96.025 of 25 items scored | 73.668.9–78.3 | 82–155 | 13.8ROCm 7.1 | 5.6 | 4–9B |
| 1 | Gemma4 E4B IT | 96.025 of 25 items scored | 95.593.0–97.7 | 2–51 | 13.2ROCm 7.1 | 7.9 | 4–9B |
| 1 | Nemotron 3.5 Bolt E4B | 96.025 of 25 items scored | 90.386.3–93.9 | 8–82 | 9.8ROCm 7.1 | 8.6 | 4–9B |
| 1 | Olmo-3-7B-Instruct | 96.025 of 25 items scored | 82.978.8–86.9 | 52–118 | 10.2ROCm 7.1 | 5.4 | 4–9B |
9–16B
Gemma 4 12B IT leads its size class with 100.0 on a scale of 0 to 100; its Overall is 98.6 (95% interval 97.0–99.8).
| Place | Model | Math (0 to 100) | Overall | Rank range | Decode (tok/s) | Memory needed (GB, est.) | Size class |
|---|---|---|---|---|---|---|---|
| 1 | Gemma 4 12B IT | 100.025 of 25 items scored | 98.697.0–99.8 | 1–12 | 8.1ROCm 7.1 | 8.9 | 9–16B |
| 2 | gemma-3-12b-it | 96.025 of 25 items scored | 92.789.7–95.4 | 8–66 | 6.1ROCm 7.1 | 9.8 | 9–16B |
| 3 | Translategemma 12B ITfine-tune of Gemma 3translation model | 92.025 of 25 items scored | 89.686.4–92.9 | 13–80 | 7.6ROCm 7.1 | 9.7 | 9–16B |
| 4 | DeepSeek-R1-Distill-Qwen-14B | 88.025 of 25 items scored | 85.381.2–89.2 | 38–106 | 7.1ROCm 7.1 | 10.8 | 9–16B |
| 4 | Ministral 3 14B Instruct (2512) | 88.025 of 25 items scored | 92.790.0–95.2 | 9–61 | 5.6ROCm 7.1 | 10.9 | 9–16B |
16–35B
gemma-4-26B-A4B-it leads its size class with 100.0 on a scale of 0 to 100; its Overall is 97.8 (95% interval 95.8–99.3).
| Place | Model | Math (0 to 100) | Overall | Rank range | Decode (tok/s) | Memory needed (GB, est.) | Size class |
|---|---|---|---|---|---|---|---|
| 1 | gemma-4-26B-A4B-it | 100.025 of 25 items scored | 97.895.8–99.3 | 1–22 | 15.7ROCm 7.1 | 21.6 | 16–35B |
| 1 | gemma-4-31B-it | 100.025 of 25 items scored | 99.398.3–100.0 | 1–5 | 3.8ROCm 7.1 | 22.4 | 16–35B |
| 1 | Muse Glimmer 30B | 100.025 of 25 items scored | 97.996.0–99.3 | 1–23 | 3.6ROCm 7.1 | 21.8 | 16–35B |
| 4 | gemma-3-27b-it | 96.025 of 25 items scored | 95.893.7–97.5 | 3–36 | 2.7ROCm 7.1 | 20.9 | 16–35B |
| 4 | gpt-oss-20b | 96.025 of 25 items scored | 95.092.2–97.6 | 2–60 | 14.9ROCm 7.1 | 16.6 | 16–35B |
> 35B
Flint 40B A2B leads its size class with 100.0 on a scale of 0 to 100; its Overall is 89.0 (95% interval 85.1–92.6).
| Place | Model | Math (0 to 100) | Overall | Rank range | Decode (tok/s) | Memory needed (GB, est.) | Size class |
|---|---|---|---|---|---|---|---|
| 1 | Flint 40B A2B | 100.025 of 25 items scored | 89.085.1–92.6 | 14–90 | 25.9ROCm 7.1 | 27.6 | > 35B |
| 2 | GLM-4.5-Air | 96.025 of 25 items scored | 92.889.7–95.6 | 6–68 | 6.0Vulkan | 87.6 | > 35B |
| 2 | gpt-oss-120b | 96.025 of 25 items scored | 98.696.8–100.0 | 1–14 | 11.0Vulkan | 78.4 | > 35B |
| 4 | Qwen3-Coder-Nextcode model | 92.025 of 25 items scored | 96.494.2–98.3 | 2–42 | 18.1Vulkan | 62.1 | > 35B |
| 5 | Sage-A1.5-Q4_K_M | 88.025 of 25 items scored | 94.591.8–96.7 | 5–56 | 22.6ROCm 7.1 | 26.4 | > 35B |
A higher place is not always a clear win. When two models have overlapping rank ranges, this battery cannot tell them apart; see how sure a score is. A dash means not measured or not scored, which is different from 0. Speeds were measured on one laptop and are comparable only between models on the same backend.
More: all the best-model pages, the hardware picker, speed and memory for every model.