DeepSeek-V4-Pro-Base
DeepSeek · 2026-04-24 · 1.6T parameters
DeepSeek-V4-Pro-Base is the pretrained base checkpoint behind DeepSeek-V4-Pro, published with the V4 preview release on April 24, 2026 in FP8 mixed precision. DeepSeek markets it at 1.6T total and 49B activated parameters; its released safetensors contain 1,600,790,440,862 stored values including the multi-token-prediction module and quantization scales. It was pretrained on more than 32T tokens with hybrid Compressed Sparse and Heavily Compressed Attention, manifold-constrained hyper-connections and the Muon optimizer, and supports a 1M-token context.
Benchmark scores
| Benchmark | Score |
|---|---|
| BBH (BIG-Bench Hard) | 87.5 |
| BIG-Bench Extra Hard | 29.8 |
| C-Eval | 93.1 |
| Capability | 131.1 |
| CLUEWSC | 85.2 |
| CMATH | 90.9 |
| CMMLU | 90.8 |
| DROP | 88.7 |
| GSM8K | 92.6 |
| HellaSwag (Unspecified Scoring Protocol) | 88.0 |
| HumanEval | 76.8 |
| LongBench v2 | 51.5 |
| MATH | 64.5 |
| MGSM | 84.4 |
| MMLU | 90.1 |
| MMLU Pro | 73.5 |
| MMLU-Redux | 90.8 |
| MMMLU | 90.3 |
| SimpleQA Verified | 55.2 |
| TriviaQA | 85.6 |
| WinoGrande | 81.5 |