MIT is on the Hugging Face card for DeepSeek-V4-Pro, not a community license with MAU gates. DeepSeek posted Pro and Flash checkpoints there. Paper arXiv:2606.19348 landed 26 April 2026. You can cite the license in an RFP. You still cannot skip export review.
The news
DeepSeek-AI posted checkpoint files for DeepSeek-V4-Pro (1.6 trillion total parameters, 49 billion activated per forward pass) and DeepSeek-V4-Flash (284 billion total, 13 billion activated) on Hugging Face. The DeepSeek-V4-Pro model card states: "This repository and the model weights are licensed under the MIT License."
Paper arXiv:2606.19348, titled "DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence," was published 26 April 2026. The abstract describes a hybrid attention architecture (Compressed Sparse Attention and Heavily Compressed Attention), one-million-token context support, and checkpoints linked from the Hugging Face collection deepseek-ai/deepseek-v4.
The MIT text on the repository grants permission to use, copy, modify, merge, publish, distribute, sublicense, and sell copies, if you keep the copyright notice and license text. It does not add monthly-active-user thresholds, derivative naming prefixes, or a separate commercial approval gate. Those terms show up in other frontier open-weight licenses. They aren't here.
DeepSeek also documents FP4 and FP8 mixed-precision weight formats and an encoding folder for message formatting rather than a bundled Jinja chat template. Whether a given hosting region or hardware SKU can run the published precision paths is a deployment question. It isn't a license question.
Who is bound
DeepSeek-AI shipped the weights and the license file. Anyone who downloads, fine-tunes, redistributes, or embeds the checkpoints inherits MIT's notice duty, plus whatever export, privacy, and sector rules apply to the use.
Cloud teams mirroring the weights, MLOps teams putting the checkpoint in an internal catalog, and ISVs shipping fine-tunes are bound. If you compare "open weights" vendors, read the Hugging Face license field. Don't read the launch adjectives.
What's new
DeepSeek-V4-Pro is among the largest open-weight MoE checkpoints published with MIT on both code and weights in Q2 2026, per the model card and the arXiv listing. Buyers who previously rejected community licenses with MAU gates or branding clauses lose those contract fights. Hardware cost, the inference stack, and compliance review don't move.
The million-token context claim in arXiv:2606.19348 shifts evaluation toward long-context harnesses and serving cost. License permissiveness doesn't tell you whether your retrieval stack still beats stuffing a corpus into context.
What it does not settle
License: MIT on repository and weights, stated on huggingface.co/deepseek-ai/DeepSeek-V4-Pro. Dates: arXiv 2606.19348 published 26 April 2026; Hugging Face collection updated in April 2026. Hardware: FP4 plus FP8 mixed formats on the model card; full-precision storage for Pro variants is multi-terabyte. Region: MIT is worldwide. U.S. and other export controls on advanced computing items and deemed-export rules for model weights still apply. The license does not waive EAR. Training-data disclosure beyond the arXiv report is UNKNOWN. DeepSeek API pricing for hosted inference is a separate invoice.
What to do now
Product counsel should store the MIT license hash next to the weight digest in the model registry. Platform teams should budget GPU memory against the published FP4 and FP8 formats instead of assuming BF16 fits the nodes you already bought.