DeepSeek opens the weights of V4-Flash-0731
A 304-billion-parameter agentic model appeared on Hugging Face under the MIT licence on the day it was announced, with published results on terminal, software engineering and cyber benchmarks.
Why it matters
The gap between what a frontier vendor serves through an API and what anyone can run locally narrowed to a single release.
The repository was created, the weights uploaded and the release committed on 31 July 2026, with a model card update the following day. The licence is MIT. The card states a size of 304 billion parameters and notes that a far smaller share is activated, without giving the number. For high and max reasoning effort levels it recommends a maximum output length of 384 thousand tokens. The published results are 82.7 on Terminal Bench 2.1, 54.4 on DeepSWE and 76.7 on Cybergym. The release supersedes the earlier V4-Flash preview and is described as having substantially enhanced agentic capabilities.