DeepSeek V4 Flash 0731 (ByteDance)

DeepSeek moved V4 Flash out of preview on July 31, 2026, with the build named DeepSeek-V4-Flash-0731 serving as the official API release of the model in public beta. According to the official changelog, the update was only re-post-trained, so the architecture and size stayed the same: a 284B-parameter mixture-of-experts with 13B active per token and a 1M-token context window, following the April 24 preview that shipped alongside the 1.6T-parameter V4 Pro. Three things changed in 0731: re-post-training rather than a new model, native Responses API support with the build specifically adapted for Codex so OpenAI-style Responses harnesses work without an adapter, and agent benchmark scores that DeepSeek says far exceed V4-Pro-Preview despite the smaller model running at roughly a third of the output price. Scope was limited to the Flash API, with the V4 Pro official release still pending.

Artificial Analysis reports that DeepSeek V4 Flash 0731 scores 50 on its Intelligence Index, a 10-point jump over the April 2026 DeepSeek V4 Flash and 6 points ahead of DeepSeek V4 Pro, putting it one point behind GPT-5.6 Luna (max, 51). The model also lands on the Pareto frontier for intelligence versus cost per task, with the analysis attributing much of its cost advantage to DeepSeek's 98% cache-hit discount on its first-party API. Architecture and size are unchanged from the earlier V4 Flash at 284B total parameters, 13B active at inference, and a 1M token context window. Agentic results improve sharply, with a GDPval-AA v2 Elo of 1559 versus 1189 previously, Terminal-Bench 2.1 up 17 points to 79%, and tau3-Bench Banking up 8 points to 31%. The AA-Omniscience Index improves to -16 via a 12-point lower hallucination rate, and full open weights were expected in coming weeks.

The independent tracker v4flash.com records that on July 31, 2026 DeepSeek released DeepSeek-V4-Flash-0731 as the official release of V4 Flash, in public beta on the API, superseding the April 24 Preview while keeping the same architecture and size and changing post-training only. The stable API model ID deepseek-v4-flash did not change, so existing API users received the new version without code changes. The site separates the model label from the service label: first-party sources call 0731 the official release but describe the API as public beta, with DeepSeek's own homepage banner using the same wording. Nothing reviewed calls it GA, so endpoints, limits, and request or response semantics may still change, including the documented default account concurrency of 2500. The same day added a native Responses API for Codex-style agents and MIT-licensed open weights on Hugging Face.