Mistral Large 4 goes open weight: a 1-trillion-parameter model you can run on your own servers

Mistral put Mistral Large 4 into public preview on 6 October 2026. On 29 October it released the weights of an improved version. Open weights means the model file itself is published, so anyone can download it and run it on their own hardware.

Mistral's flagship model has 1 trillion parameters, the learned values that make up a model, and reads text and images natively. Mistral says three more weeks of training since the preview improved its coding and its defensive cybersecurity work. It also says the model now reasons with fewer tokens, which lowers running costs.

The model comes under the Mistral Ecosystem Open-Weight license, which Mistral describes as a modified MIT license. It is free to use, modify, fine-tune and sell on. The one condition applies to a business that runs a Model-as-a-Service offering. It needs a separate commercial licence from Mistral before it uses the model commercially.

Mistral says the model was trained from scratch on European infrastructure. The model card puts that at 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own datacentres in Europe.

On Mistral's stated terms, a bank, insurer or public body can run the model for its own work without paying a licence fee.

The condition applies to anyone who sells access to the model itself. Model-as-a-Service means offering the model to others as a hosted service, such as an API. Software firms that plan to put the model behind their own product should read the licence text closely. Mistral hasn't published the terms of the separate licence.

A buyer with sovereignty rules gets a model trained in Europe that can run without Mistral's API. Mistral says organisations can inspect, customise and deploy it on their own infrastructure.

On Mistral's API, the model costs $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. A token is a word or part of a word. Cached input is repeated text, such as the same long instructions sent with every request. By our arithmetic it costs a tenth of fresh input.

Mistral offers the model in Mistral Studio, on its API as mistral-large-4, in Vibe Code and on OpenRouter. The weights are on Hugging Face, ready to serve with the open-source tools vLLM or SGLang. NVIDIA offers hosted endpoints and a ready-made NIM container.

Microsoft Foundry has the model too, with private deployment options for Azure developers. In Mistral's announcement, Microsoft says it is the first cloud provider to bring the model to customers. Neither side gives a Foundry price or a list of regions.

Mistral says customers can adapt the model to their own data with Mistral Forge, and build and govern agents in Mistral Studio. Its Applied AI experts can work alongside a customer's team.

Take a team that sends 100 million input tokens and gets back 20 million output tokens a month. By our arithmetic, it would pay $70 plus $42, or $112, at Mistral's list price before any caching. Price your own monthly volume the same way, from your current bills.

Of the three routes, the API is likely the cheapest to start and needs no hardware. A private deployment on Foundry runs the model within Azure. Running the weights on your own servers gives the most control, and it also takes the most hardware and the most people.

The pages we read don't say where Mistral's API processes data. A buyer with residency rules should ask Mistral where processing runs before sending such data to the API, or look at Foundry or self-hosting.

Mistral Large 4 is a mixture of experts. The model is split into many specialist parts, called experts, and only a few of them run for each token. It has 128 experts, of which 4 run for each token, plus 1 shared expert.

The model card says 49 billion parameters run for each token, or 52 billion counting its embedding and output layers. By our arithmetic, that is about one in twenty of the full trillion. The model card also offers a smaller NVFP4 version for NVIDIA Blackwell GPUs, which Mistral says uses less memory and runs faster with the same performance.

The model card's serving example spreads the model across a node of 8 GPUs, and the card says to adjust that to your own setup. It handles up to 1 million tokens of context, which is the amount of text it can consider at once.

Self-hosting a model of this size means buying or renting the GPUs. Someone also has to keep the model served and patched, and respond when it fails.

Only about one in twenty parameters works on each token, which can keep the computing cost of each token down. The whole model still has to be loaded into GPU memory, which is likely why the example uses a full node of 8 GPUs.

Self-hosting is likely to make sense only where a company's data can't leave its own servers. Volume is the other reason, once owned hardware costs less than paying per token.