Mistral Large 4 open weights: can you run it on your own servers?

TL;DR: on October 29, 2026, Mistral AI published the weights of Mistral Large 4, its flagship model, on Hugging Face. A company can now host it on its own servers without sending data to an API. It is possible, but heavy: you need four to eight latest-generation GPUs, and the licence requires a commercial agreement as soon as you sell access to the model to third parties.

What was released

• Repository: mistralai/Mistral-Large-4-1T-A52B, plus a quantized -NVFP4 variant, ungated.
• Size: about 1.05 trillion parameters, 52 billion active per token (mixture of experts: 128 experts, 4 active, plus one shared expert).
• Capabilities: text and images in, text out, 1M-token context announced, adjustable reasoning effort.
• Ranking: according to Artificial Analysis, the best open-weights model developed outside China.

Hardware is the real topic

The weights are about 1.06 TB in bf16 and 610 GB in NVFP4. The published vLLM recipe uses eight GPUs with tensor parallelism. The NVFP4 version runs, according to Mistral, on four Nvidia B200 or B300 GPUs (Blackwell). This is an infrastructure project, to compare with the cost of the API.

The licence: free, with one exception

The Mistral Ecosystem Open-Weight License is a modified MIT licence. Internal use, commercial use and embedding the model in a product are free, with no revenue threshold. Selling access to the model as a service (an API or inference for third parties who control the requests) requires a separate commercial licence with Mistral, affiliates included. This is not legal advice.

Before putting Large 4 in production

1. Match the model to the need: many agent use cases run fine on smaller models (Mistral Small, Ministral).
2. Size and serve it: GPUs, KV cache memory, autoscaling, monitoring. See the four questions for serving an open LLM with vLLM.
3. Make everything work offline: on an air-gapped server, every component must find its files locally. I fixed a bug in mistral-common that prevented loading a tokenizer already on disk (article, fix merged by Mistral AI).
4. Evaluate on your own cases: a set of real questions with expected answers, replayed against Large 4, a smaller model and the API.
5. Add guardrails back: when self-hosting, the Mistral API protections do not apply.

Takeaway

Sovereign AI is becoming a credible option for organisations that cannot send their data to an API. But downloading the weights is only the start: hardware cost, licence and industrialisation decide whether the project reaches production.

Sources: Hugging Face model card · OrcaRouter: weights and licence · French version

---

Zakaria Khchiche, Data & AI Tech Lead (freelance, Paris), 200+ AI agents in production, contributor to Mistral AI and Microsoft Agent Framework. LinkedIn · Website