GSM8K — Exact-Match protocol, diff from accuracy-based gsm8k/GSM8K; DeepSeek-V4-family source.
| # | Model | Vendor | Best score | Runs | Last seen |
|---|---|---|---|---|---|
| 1 | DeepSeek-V4-Pro-Base | DeepSeek | 92.6 | 4 | 2026-06-27 |
| 2 | DeepSeek-V3.2-Base | DeepSeek | 91.1 | 4 | 2026-06-27 |
| 3 | DeepSeek-V4-Flash-Base | DeepSeek | 90.8 | 4 | 2026-06-27 |
| 4 | DeepSeek-V3 | DeepSeek | 89.3 | 1 | 2026-05-01 |
| 5 | Qwen2.5 Instruct 72B | Alibaba | 88.3 | 1 | 2026-05-01 |
| 6 | Llama 3.1 Instruct 405B | Meta | 83.5 | 1 | 2026-05-01 |
| 7 | DeepSeek-V2 | DeepSeek | 81.6 | 1 | 2026-05-01 |
| 8 | Qwen2.5 7B Base | Alibaba | 36.4 | 1 | 2026-05-03 |
| 9 | Qwen2.5 3B Base | Alibaba | 29.2 | 1 | 2026-05-03 |
| 10 | Qwen3 4B Base Dense | Alibaba | 28.3 | 1 | 2026-05-03 |
| 11 | Gemma 3 1B Base Dense | 9.7 | 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.