BIG-Bench Hard — DeepSeek base-model reporting w/ explicit EM metric; kept apart from metric-unspecified bbh
| # | Model | Vendor | Best score | Runs | Last seen |
|---|---|---|---|---|---|
| 1 | DeepSeek-V3.2-Base | DeepSeek | 87.6 | 4 | 2026-06-27 |
| 2 | DeepSeek-V3 | DeepSeek | 87.5 | 1 | 2026-07-06 |
| 3 | DeepSeek-V4-Pro-Base | DeepSeek | 87.5 | 4 | 2026-06-27 |
| 4 | DeepSeek-V4-Flash-Base | DeepSeek | 86.9 | 4 | 2026-06-27 |
| 5 | Llama 3.1 Instruct 405B | Meta | 82.9 | 1 | 2026-07-06 |
| 6 | Qwen2.5 Instruct 72B | Alibaba | 79.8 | 1 | 2026-07-06 |
| 7 | DeepSeek-V2 | DeepSeek | 78.8 | 1 | 2026-07-06 |
Best tracked score per model (default configuration; source-attributed and verification-tiered). Open a model for its full benchmark surface, provenance and pricing.