DeepSeek ships V4-Flash-0731 under MIT — a production agent model whose results beat its bigger sibling
DeepSeek's production release of V4-Flash — 284B total parameters, 13B active, 1M-token context — lands under the MIT license with agent benchmarks the company says far exceed its larger V4-Pro-Preview. Post-training alone produced the jump, and the weights are free to use commercially.
The headline is what changed: nothing about the architecture, everything about the post-training. V4-Flash-0731 is the April preview retrained through an improved post-training pipeline, and the result — Terminal Bench 2.1 at 82.7, agentic scores above the bigger V4-Pro-Preview — is a clean demonstration that post-training is now where capability comes from, not parameter count.
The MIT license is the strategic part. A production-grade, million-token-context agent model with 13 billion active parameters, free for commercial use with no strings, resets the floor for what anyone can charge for comparable capability. It also ships with native Responses API and Codex support — open weights arriving pre-integrated into the very harnesses the closed labs built.
Existing users get the upgrade silently: applications calling deepseek-v4-flash pick up the new build without changing a line. That deployment model — continuous capability delivery under an immutable open license — is the open-weight ecosystem behaving like a cloud service, and it is why the frontier labs' pricing pressure is not letting up.
MarkTechPost — DeepSeek upgrades V4-Flash-0731 with major agentic and coding gains → · Simon Willison — deepseek-ai/DeepSeek-V4-Flash-0731 → · Open Source For You — DeepSeek open sources production V4-Flash under MIT licence →