DeepSeek released the official version of DeepSeek-V4-Flash-0731 on July 31, moving the model out of preview and into public beta on its API, according to the company’s own release notes. The open-weight model, published on Hugging Face under an MIT license, keeps the same architecture as the preview build — a mixture-of-experts network with 256 routed experts plus one shared expert, of which only six activate per token — but was retrained with a stronger post-training pipeline aimed at agentic coding and tool-use tasks.
Beats its own bigger model
DeepSeek’s benchmark data shows the retrained Flash model now outperforms DeepSeek-V4-Pro (Preview), the larger model in the same family, on several coding and agent tests, including Terminal-Bench 2.1 (82.7), Cybergym (76.7) and DeepSWE (54.4) — despite activating only a fraction of Pro’s parameters per token. The model also supports a one-million-token context window and now natively handles OpenAI’s Responses API format, which DeepSeek says it adapted specifically for use inside Codex-style coding agents.
Pricing is unchanged from the preview release: $0.14 per million input tokens on a cache miss, $0.0028 per million on a cache hit, and $0.28 per million output tokens — roughly a third of what DeepSeek charges for V4-Pro. The company’s documentation notes that pricing will later shift to a peak/off-peak schedule, doubling during set daytime hours in China.
Part of a broader efficiency race
The release continues a pattern among open-weight Chinese labs of competing on efficiency rather than raw scale: Moonshot AI’s Kimi K3 and Ant Group’s Ling-3.0-Flash both launched recently as leaner counterparts to their makers’ largest models. DeepSeek itself became shorthand for that shift after earlier releases unsettled assumptions in Silicon Valley about the cost of building frontier AI.
Developers already calling the deepseek-v4-flash endpoint are upgraded to the new build automatically, without code changes, DeepSeek said. The DeepSeek-V4-Pro API and the company’s consumer app and web products are unaffected by the update.