DeepSeek V4 Official Paper Released

The official paper for DeepSeek V4 has been released, revealing more concrete training and serving designs than the 58-page preview.

• FP4 QAT was applied directly during the late training stage, quantizing MoE expert weights to FP4, and FP4 activations were also used in the QK path of the CSA indexer.
• The QK selector maintained a 2x speedup and 99.7% recall, and inference is also explained to be performed directly on FP4 weights.
• At 1M context, efficiency figures show V4-Pro at 27% FLOPs and 10% KV cache, while V4-Flash is at 10% FLOPs and 7% KV cache.
• To address loss spikes in large-scale MoE, two mechanisms were added: anticipatory routing and SwiGLU clamping. The former delays some router updates to break the feedback loop of anomalous signals, while the latter limits the output of the linear/gate path to reduce divergence. Anticipatory routing has an overhead of about 20%, but it is only activated during loss spikes.
• For reward training, instead of a separate reward model, a generative reward model was configured where the same model handles both generation and evaluation, reducing manual labeling and training evaluation that includes reasoning.
• In released human evaluations, the model achieved a 62.7% win rate in Chinese writing, 77.5% in writing quality, and in white-collar tasks, V4-Pro-Max outperformed Opus 4.6 Max with a 63% non-loss rate.
• In coding agent evals, 52% said it could be used as a primary code model, 39% responded positively, and under 9% responded negatively.