DeepSeek-V4-Flash-0731
Official release of DeepSeek-V4-Flash, with stronger agentic skills and open weights under the MIT License.
DeepSeek-V4-Flash-0731 is DeepSeek's official release of DeepSeek-V4-Flash, replacing the preview with substantially enhanced agentic capabilities, and its weights are licensed under the MIT License on Hugging Face. DeepSeek's own API has retired it and routes the old name to V4.1-Flash.
About DeepSeek-V4-Flash-0731
DeepSeek's model card says 0731 has the same model structure as DeepSeek-V4-Flash-DSpark, so it comes with a speculative decoding module attached. A reasoning_effort parameter takes three levels, low, high and max, which set how much the model deliberates before answering.
In DeepSeek's table, 0731 beats DeepSeek-V4-Pro (Preview) on all nine listed benchmarks despite a far smaller activated parameter count, scoring 82.7 on Terminal Bench 2.1, 54.4 on DeepSWE and 70.3 on Toolathlon-Verified. Opus-4.8 posts 85.0, 58.0 and 76.2 on those three, while GLM-5.2 scores below 0731 on each of the eight rows where it has a result.
The card's vLLM and SGLang commands switch on DSpark speculative decoding, and its vLLM example serves the model on a single 4×GB300 node; an inference folder covers running DeepSeek-V4 locally, with weight conversion and chat demos. DeepSeek recommends temperature 1.0, top_p 0.95 for agentic use and 1.0 otherwise, and a maximum output length of 384K tokens at the high and max effort levels.
Sources: huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731, deepseek.com/en/news/deepseek-v4-1-flash and api-docs.deepseek.com/quick_start/pricing, read 2026-10-26
DeepSeek-V4-Flash-0731 pros and cons
Pros
- The repository and the model weights are released under the MIT License.
- Against the V4-Flash preview, DeepSeek's table shows large gains, such as 54.4 against 7.3 on DeepSWE and 82.7 against 61.8 on Terminal Bench 2.1.
- DSpark speculative decoding is built in and is switched on with a single vLLM flag.
Cons
- DeepSeek's own API has retired V4-Flash: the old model name now routes to V4.1-Flash.
- In DeepSeek's own table, Opus-4.8 scores higher than 0731 on all nine benchmarks.
- Two of the nine benchmarks, DSBench-FullStack and DSBench-Hard, are DeepSeek's internal test sets.
- There is no Jinja-format chat template: prompts are built with Python scripts from the release's encoding folder.
Tagged
- Web-based
- Agentic Mode
DeepSeek-V4-Flash-0731 alternatives
The closest alternatives to DeepSeek-V4-Flash-0731.
Related reading
Featured on Vibedonalds
Own DeepSeek-V4-Flash-0731? Add this badge to your site to show you’re listed — and link back to your profile here.
<a href="https://vibedonalds.com/tools/deepseek-v4-flash-0731" target="_blank" rel="noopener">
<img src="https://vibedonalds.com/badge/featured-on-vibedonalds.svg" alt="DeepSeek-V4-Flash-0731 — Featured on Vibedonalds" width="240" height="60" loading="lazy" />
</a>Frequently asked questions
- What is DeepSeek-V4-Flash-0731?
- DeepSeek-V4-Flash-0731 is an official release of DeepSeek-V4-Flash, with stronger agentic skills and open weights under the MIT License.
- What are the alternatives to DeepSeek-V4-Flash-0731?
- The closest alternatives to DeepSeek-V4-Flash-0731 are Beam, Hy4 preview.
- What are the downsides of DeepSeek-V4-Flash-0731?
- DeepSeek's own API has retired V4-Flash: the old model name now routes to V4.1-Flash. In DeepSeek's own table, Opus-4.8 scores higher than 0731 on all nine benchmarks. Two of the nine benchmarks, DSBench-FullStack and DSBench-Hard, are DeepSeek's internal test sets. There is no Jinja-format chat template: prompts are built with Python scripts from the release's encoding folder.
- Is DeepSeek-V4-Flash-0731 free to use?
- Its repository and model weights are licensed under the MIT License, and the weights are published on Hugging Face. The model card does not state a price for any hosted API.
- Is there a DeepSeek-V4-Flash-0731 API?
- Not from DeepSeek any more: its pricing page says requests to the legacy name deepseek-v4-flash are served by the DeepSeek-V4.1-Flash-1017 model and billed at the Flash price. The model card shows how to run your own OpenAI-compatible server with vLLM or SGLang, and the Hugging Face page lists Novita and two more inference providers for the model.
- Is there a GGUF of DeepSeek-V4-Flash-0731?
- DeepSeek's model card does not mention GGUF; the release is tagged FP8 and 8-bit precision. The Hugging Face model tree lists 202 quantizations of it, which the page points to for llama.cpp, Ollama, LM Studio or any compatible app.
- How does 0731 compare with the DeepSeek-V4-Flash preview?
- DeepSeek calls 0731 the official release of DeepSeek-V4-Flash, superseding the preview with substantially enhanced agentic capabilities. In its table, 0731 scores 70.3 on Toolathlon-Verified against the preview's 49.7, and 54.2 on NL2Repo against 39.4.
- How do I run DeepSeek-V4-Flash-0731?
- The model card gives vLLM and SGLang commands that enable DSpark speculative decoding, with a vLLM example on a single 4×GB300 node, plus an inference folder for running it locally. DeepSeek recommends temperature 1.0 with top_p 0.95 for agentic use.