DeepSeek-V4-Flash-0731 (official)
Official V4-Flash, with agent benchmarks up sharply versus the preview (DeepSWE 12.8 to 54.4, Terminal Bench 2.1 72.1 to 82.7) and a native Responses API for Codex.
Post-training upgrade focused on agentic work (low/high/max reasoning effort), built-in DSpark speculative decoding, and OpenAI Responses API compatibility so it plugs into Codex-style clients. MIT license; Hugging Face lists 304B parameters.
- Date
- Friday, 31 July 2026
- Lab
- DeepSeek
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| DeepSWE | 54.4 (preview 12.8) Terminal Bench 2.1 82.7 vs 72.1; Toolathlon-Verified 70.3 vs 55.9; NL2Repo 54.2 vs 38.5 | company |
The API changelog labels this a public beta. The 304B figure differs from the preview's 284B on the Hugging Face cards; the cause (extra modules such as MTP/DSpark) is not stated.
Sources
- api-docs.deepseek.com/updates
- huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731
- api-docs.deepseek.com/news/news260731
This record was checked against its sources on 6 October 2026. How we check