DeepSeek-V4-Flash-0731 · DeepSeek · Model History

Claims and evidence

Released 31 July 2026Released in the DeepSeek API as a public beta under the unchanged model name deepseek-v4-flash; the V4-Pro API and the app and web models were not changed by this update. Primary source[2]Date: 2026-07-31

Status AvailableWeights remain downloadable from the deepseek-ai Hugging Face organisation. In the DeepSeek API the model was retired on 2026-09-10; the legacy name deepseek-v4-flash now routes to DeepSeek-V4.1-Flash. Primary source[1]License [2]Date: 2026-09-10
Replaced by DeepSeek-V4.1-FlashIn the DeepSeek API: V4-Flash was retired on 2026-09-10 and deepseek-v4-flash temporarily routes to DeepSeek-V4.1-Flash. Primary source[2]Date: 2026-09-10 [3] [4]footnote 1

Change · Architecture Ships with an attached DSpark speculative decoding module; DeepSeek says architecture and size are otherwise unchanged and only post-training was redone.Compared with DeepSeek-V4-Flash Primary source[1]Introduction [2]Date: 2026-07-31
Change · Reasoning Reasoning effort is set through a reasoning_effort parameter with three levels: low, high and max.Compared with DeepSeek-V4-Flash Primary source[1]Chat Template
Change · Tool use Natively supports the OpenAI Responses API format and is adapted for Codex.Compared with DeepSeek-V4-Flash Primary source[2]
Change · Other DeepSeek reports substantially stronger agentic capabilities than the preview.Compared with DeepSeek-V4-Flash Primary source[1]Introduction

Feature Reasoning modeReasoning effort levels low, high and max. Primary source[1]Chat Template [2]Date: 2026-08-13
Open weights YesWeights under the MIT License. Primary source[1]License
Access API, Open-weights download Primary source[2]Date: 2026-07-31 [1]License

No known predecessor.

None recorded.

AI Radar coverage starts in June 2026; no coverage linked yet.