DeepSeek V4 Flash 0731 goes open weights under MIT, agentic leap
Same 284B/13B MoE and $0.14 pricing as the preview, but Terminal Bench leaps to 82.7 and it lands top-3 among open-weights models. Day-0 GGUFs are up.
Copy markdown
MIT weights, self-host today
DeepSeek pushed the full V4 Flash 0731 weights to Hugging Face under the MIT license — unrestricted commercial use, self-hostable via vLLM or SGLang. It's the official release that supersedes July's preview, so you can run the exact model behind the API on your own boxes.
Agent scores nearly doubled
It's a re-post-train, not a new architecture, aimed squarely at agents: Terminal Bench 2.1 jumps from 61.8 to 82.7 and DeepSWE from 7.3 to 54.4. The Artificial Analysis Intelligence Index climbed 40 to 50, putting it among the top 3 open-weights models.
Still 13B active, still $0.14
The refresh keeps the 284B-total / 13B-active MoE and 1M-token context, and API pricing holds at $0.14/$0.28 per million tokens with cache hits at $0.003 (~98% off). A speculative-decoding module ships attached for faster local inference.
Build this weekend: an agent you fully own
Unsloth's day-0 GGUFs are already live, so a quantized 13B-active model with 1M context fits on a single workstation. Native Responses API support plus low/high/max reasoning make it a drop-in backend for a coding or terminal agent — no vendor in the loop.
The catch: verbose, and pricier than it looks
Flash is chatty — it spent ~230M output tokens running the Intelligence Index versus a ~92M median, so real task bills land above the headline rate. DeepSeek also has peak-hour pricing (all rates double for 7 hours a day) planned, though not yet switched on.