DeepSeek-V4-Flash-0731: The Agentic Coding Upgrade That Closes the Gap With Pro
DeepSeek shipped V4-Flash-0731 on July 31, an updated build of its existing Flash model that trades a general-purpose post-training run for one targeting specifically agentic and coding performance. The architecture hasn't changed — it's still the 284B total / 13B active parameter Mixture-of-Experts with a 1M-token context window — but the post-training focus shows clearly in the benchmark spread.
TerminalBench 82.7 and NL2Repo 54.2 are the two numbers DeepSeek is leading with. The former measures long-horizon terminal agent tasks; the latter evaluates code generation at the repository level, where a model must understand file relationships and cross-module dependencies rather than just isolated functions. Both place V4-Flash-0731 above the previous Flash version and close to Pro-tier models on coding-specific evals.
The Artificial Analysis Intelligence Index puts the updated model at 50 — a 10-point improvement over the previous Flash version — closing to within reach of several established Pro-tier offerings that cluster around the 55–60 range.
The update also ships a speculative decoding module matching the structure used in DeepSeek's DSpark variant, improving inference throughput without changing output quality. The API entered public beta on July 31 alongside the update, with full support for the Responses API format that coding agents typically rely on.
Critically, the MIT license carries over: weights are publicly downloadable, and the community has moved quickly. Unsloth published optimised GGUF quantisations ranging from 3-bit to lossless 8-bit within hours of the weight release — meaning the model is running locally on consumer and prosumer hardware the same day it launched.
The V4-Flash-0731 update only affects the Flash API endpoint. V4-Pro and the main DeepSeek chat interface remain unchanged.
This update doesn't come with the broader narrative framing of a new model launch, but for developers running coding agents at scale, a 10-point intelligence index jump with open weights and a Responses-compatible API is the kind of incremental release that changes which model is in production by end of week.
