DeepSeek publishes the 1.7T-parameter DeepSeek-V4-Pro-0813 weights under the MIT licence

Sources: primary, discussion, discussion
Summary: The Hugging Face repository was created 2026-08-13 and states that the repository and the model weights are licensed under the MIT License, with no revenue or user thresholds attached. The card lists 1.7T parameters, a DSpark speculative decoding module attached to the DeepSeek-V4-Pro preview structure, reasoning_effort levels of low, high, and max, and launch flags for vLLM and SGLang. The release ships no Jinja chat template and provides a Python encoding folder instead, so an existing serving stack needs a change before it can run the model, and the benchmark table on the card is DeepSeek's own and was not reproduced here.
Why it matters: A 1.7T-parameter model published under a permissive licence with no usage thresholds makes self-hosting a hardware decision rather than a vendor contract decision.