MATH (Hendrycks et al.) — DeepSeek-V4-family report, explicit EM axis; distinct from generic math aggregate bucket.
| # | Model | Vendor | Best score | Runs | Last seen |
|---|---|---|---|---|---|
| 1 | Qwen2.5 7B Base | Alibaba | 85.4 | 1 | 2026-05-03 |
| 2 | Qwen3 4B Base Dense | Alibaba | 75.4 | 1 | 2026-05-03 |
| 3 | DeepSeek-V4-Pro-Base | DeepSeek | 64.5 | 4 | 2026-06-27 |
| 4 | DeepSeek-V3 | DeepSeek | 61.6 | 1 | 2026-07-06 |
| 5 | DeepSeek-V3.2-Base | DeepSeek | 60.5 | 4 | 2026-06-27 |
| 6 | DeepSeek-V4-Flash-Base | DeepSeek | 57.4 | 4 | 2026-06-27 |
| 7 | Qwen2.5 Instruct 72B | Alibaba | 54.4 | 1 | 2026-07-06 |
| 8 | Llama 3.1 Instruct 405B | Meta | 49 | 1 | 2026-07-06 |
| 9 | DeepSeek-V2 | DeepSeek | 43.4 | 1 | 2026-07-06 |
| 10 | Qwen2.5 3B Base | Alibaba | 17.7 | 1 | 2026-05-03 |
| 11 | Gemma 3 1B Base Dense | 7.2 | 1 | 2026-05-03 |
Best tracked score per model (default configuration; source-attributed and verification-tiered). Open a model for its full benchmark surface, provenance and pricing.