The Frontier · Intelligence · Entity
DeepSeek-V4-Pro
model · Model capabilities, AI infrastructure
| Metric | Value | Unit | As of | Origin | Verification | Context | Source | Evidence | Article |
|---|---|---|---|---|---|---|---|---|---|
| CorpusQA 1M accuracy | 62 | accuracy | — | paper | Reported (unverified) | CorpusQA 1M — Think Max | arXiv — DeepSeek-V4 (2606.19348) | Table 6 — CorpusQA 1M ACC | million-token-mrcr-claim-check |
| MRCR 1M mean match ratio | 83.5 | mmr_score | — | paper | Reported (unverified) | MRCR 1M — Think Max / Max mode | arXiv — DeepSeek-V4 (2606.19348) | Table 6 — MRCR 1M MMR | million-token-mrcr-claim-check |
| MRCR 1M mean match ratio | 44.7 | mmr_score | — | paper | Reported (unverified) | MRCR 1M — Non-Think mode | arXiv — DeepSeek-V4 (2606.19348) | Table 7 — Non-Think mode | million-token-mrcr-claim-check |
| Relative single-token inference FLOPs vs baseline | 0.27 | ratio | — | paper | Reported (unverified) | DeepSeek-V4 efficiency at filled 1M context vs V3.2 | arXiv — DeepSeek-V4 (2606.19348) | Abstract / Figure 1 efficiency claim | deepseek-v4-attention-efficiency |
| Relative accumulated KV cache vs baseline | 0.1 | ratio | — | paper | Reported (unverified) | DeepSeek-V4 efficiency at filled 1M context vs V3.2 | arXiv — DeepSeek-V4 (2606.19348) | Abstract / Figure 1 efficiency claim | deepseek-v4-attention-efficiency |