Open weights under the MIT license: what DeepSeek is releasing

Licenses are a major factor in the current model race. The DeepSeek-V4.1-Flash model card states it plainly: the repository and model weights are covered by the MIT license. It is one of the most permissive open source licenses, requiring no disclosure of changes and imposing no restrictions on commercial use.

The release is larger than it might seem. The Hugging Face repository holds 48 weight files in safetensors format, alongside a set of tools. There is no Jinja chat template, but there is standalone reference encoding code in Python. It handles multi-turn conversations, tool calls, reasoning mode, numerical reasoning effort, and images woven into the text stream.

Not just the model files

The interoperability statement has value in itself. The same model card lists deployment variants for vLLM, SGLang, TensorRT, and Docker Model Runner, plus FP8, BF16, GPTQ, AWG, and GGUF quantizations. It also names outside inference providers, including novita, fireworks, featherless, and deepinfra, that run the model on their own hardware. The model also runs in llama.cpp, Ollama, and LM Studio. The weights are not just released on paper; they can genuinely be run outside the maker's infrastructure.

DeepSeek took an earlier step in the same direction on 31 August 2026. The Chinese outlet IT之家 noted that DeepSeek-V4-Flash-Vision-Exp, the first multimodal model in the V4 series, available in the API since 21 August 2026, also landed on Hugging Face under the MIT license. The company opened not only the model files and tokenizer, but also the reference prompt encoding implementation and a minimal inference implementation in PyTorch. This covers the vision encoder, the aligner, DFlash attention, MoE, hyper-connections, and DSpark. The model accepts JPEG, PNG, GIF, and WebP images; the maker labelled it an experimental version.

Openness as a commercial weapon

This is not a selfless gesture. The Italian outlet CorCom argued in an analysis dated 7 September 2026 that cheap open-weight models from China can squeeze the margins of companies building the most expensive AI systems. The authors, Filippo Pallotti and Michael Strobaek of Banque Lombard Odier, add a caveat: the advantage of Chinese models is overstated. Distillation and the cost and time of completing a task can raise the real cost of work even if the price per token looks excellent. That is an important addition to the enthusiasm around the MIT license: free weights do not mean a free deployment.

Yet the figure of 621 396 downloads in the past month on the model card shows that the availability of the weights alone draws engineers in, regardless of the argument over margins.