DeepSeek has launched DeepSeek-V4-Flash-0731, a public beta available through DeepSeek’s API, with open weights published on Hugging Face under the MIT license the same day. The model uses the same architecture as the preview release, with 284 billion total parameters and 13 billion activated parameters per token, much smaller than V4-Pro’s 1.6 trillion total parameters and 49 billion activated parameters. DeepSeek attributes the performance gains to additional post-training rather than a larger model. The updated Flash version now beats the earlier V4-Pro preview on several agent-focused benchmarks, with the company reporting 82.7 on Terminal-Bench 2.1, 54.4 on DeepSWE, and 70.3 on Toolathlon-Verified. However, early independent testing by Artificial Analysis found a lower Terminal-Bench 2.1 score of 79%, suggesting reported numbers may not always match independent results. DeepSeek also shared results from several internal tests that have not yet been independently verified. The release of production-ready weights under a permissive license gives organizations more control over deployment and customization.