DeepSeek-V4-Pro-0813 GA release with DSpark speculative decoding
DeepSeek released DeepSeek-V4-Pro-0813 on Hugging Face under MIT license. It supersedes the preview version, adds a DSpark speculative decoding module, and supports reasoning effort levels low, high, and max. Reported benchmark scores include HLE 42.7 (without tools) / 60.0 (with tools), Terminal Bench 2.1 87.9, NL2Repo 61.5, Cybergym 83.3, DeepSWE 62.7, Toolathlon-Verified 74.1, Agents' Last Exam 25.7, AutomationBench (Public) 31.8, DSBench-FullStack 71.1, DSBench-Hard 67.2.
China context
- Original name
- 深度求索
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Builders outside China can download the MIT-licensed weights from Hugging Face and integrate DSpark speculative decoding using vLLM or SGLang for faster inference.
- For investors
- Investors can monitor adoption of DeepSeek-V4-Pro-0813 in open-source projects and enterprise deployments as a signal of competitive pressure on proprietary model providers.
Translated from Chinese. Quotes and facts link to the original sources.
DeepSeek released DeepSeek-V4-Pro-0813, the official GA version of DeepSeek-V4-Pro, on Hugging Face under MIT license. It includes a DSpark speculative decoding module and supports three reasoning effort levels. The model card reports benchmark improvements over the preview version and competitive performance with proprietary models.
DSpark speculative decoding is enabled via a single flag in vLLM (--speculative-config '{"method":"dspark","num speculative tokens":7,"draft sample method":"greedy"}') and SGLang (--speculative-algorithm DSPARK), with target and draft weights from the same checkpoint. The model does not include a Jinja chat template; instead, a dedicated encoding folder provides Python scripts for OpenAI-compatible message encoding. Recommended sampling parameters are temperature=1.0, top_p=0.95 for agentic scenarios, and max output length 384K tokens for high/max reasoning effort.
DeepSeek continues to release open-weights models under MIT license, competing directly with proprietary models like GLM-5.2, Kimi K3, Opus-4.8, and Fable-5. The inclusion of speculative decoding in the official release signals a focus on production inference efficiency.
For developers, the MIT license and open weights allow self-hosting and fine-tuning. The DSpark speculative decoding can reduce inference latency and cost in production deployments. The model's agentic benchmark improvements may make it suitable for coding and automation tasks.
Observable next signals include adoption of the DSpark speculative decoding method in other open-source inference frameworks, community benchmarks reproducing the reported scores, and API availability of the GA model on DeepSeek's platform.