DeepSeek released DeepSeek-V4-Flash-0731 on Hugging Face under MIT license
DeepSeek released DeepSeek-V4-Flash-0731 on Hugging Face. It is a text-generation model under MIT license, using the transformers library. The model card reports benchmark scores: Terminal Bench 2.1 82.7, NL2Repo 54.2, Cybergym 76.7, DeepSWE 54.4, Toolathlon-Verified 70.3, Agents' Last Exam 25.2, AutomationBench Public 25.1, DSBench-FullStack 68.7, DSBench-Hard 59.6. It supports reasoning effort levels low, high, and max, and recommends max output length of 384K tokens for high and max effort.
China context
- Original name
- 深度求索
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can download the MIT-licensed weights from Hugging Face and run inference with vLLM or SGLang using DSpark speculative decoding.
- For investors
- The release of a high-performance open-weights model under MIT license may pressure proprietary model pricing and accelerate adoption of DeepSeek's architecture.
DeepSeek released DeepSeek-V4-Flash-0731 on Hugging Face, an official release superseding the preview version with enhanced agentic capabilities. The model has the same structure as DeepSeek-V4-Flash-DSpark, including a speculative decoding module. It is licensed under MIT and supports vLLM and SGLang inference with DSpark speculative decoding.
The model uses DSpark speculative decoding, enabled via vLLM with --speculative-config '{"method":"dspark","num speculative tokens":7,"draft sample method":"greedy"}' or SGLang with --speculative-algorithm DSPARK. It does not include a Jinja chat template; instead, a dedicated encoding folder provides Python scripts for OpenAI-compatible message encoding. Reasoning effort supports three levels: low, high, max.
DeepSeek-V4-Flash-0731 outperforms DeepSeek-V4-Pro (Preview) on listed benchmarks despite a smaller activated parameter count, and is broadly competitive with proprietary models like GLM-5.2 and Opus-4.8. The MIT license and Hugging Face release continue DeepSeek's open-weights strategy.
The MIT license allows commercial use and modification, enabling developers to integrate the model into products without licensing fees. The speculative decoding support can reduce inference latency and cost.
Observable next signals include the release of DeepSeek Harness (minimal mode) mentioned in the model card, and independent benchmark evaluations of the reported scores.