DeepSeek's V4-Flash 0731 update lifts agentic coding scores sharply

DeepSeek published DeepSeek-V4-Flash-0731 on 31 July 2026, a re-post-trained build of the V4-Flash model it first previewed in April. The architecture is unchanged — a sparse mixture-of-experts with 284 billion total parameters and roughly 13 billion active per token, a one-million-token context window and text-only input — but the tuning is aimed squarely at agentic work and software engineering. The reported gains are large. On Terminal-Bench 2.1 the model moves from 61.8 to 82.7, and on DeepSWE from 7.3 to 54.4. Independent testing by Artificial Analysis places the build at 50 points on its Intelligence Index, ten points above the previous Flash release and level with Gemini 3.6 Flash. Weights are available on Hugging Face under the MIT licence. The hosted API is priced at 0.14 US dollars per million input tokens and 0.28 dollars per million output tokens, with cached input billed at a small fraction of that — pricing that undercuts most models tested at a comparable capability level.