AI models

Mistral Large 4 is now open weight: Europe's trillion-parameter model is ready to download

On October 29, 2026 Mistral AI published the weights of Mistral Large 4, nicknamed "Le Chonk", on Hugging Face: a multimodal model of about one trillion parameters, trained in Europe, that anyone can download and run on their own servers. Mistral Large 4 had been an API-only preview since October 6. Here is what the release includes, what the license says, what the API costs and what it changes for you.

Published on 7 min readBy the editorial team of The AI Creator System

Leggi questo articolo in italiano

AI-generated cover: a large block of cyan and violet crystal cubes opening like a treasure chest, with the title Mistral Large 4 now open weight
Cover image generated with artificial intelligence. How we use AI

Mistral Large 4 open weights: what was released on October 29

Mistral Large 4 is Mistral AI's largest model, and since October 29, 2026 its weights are free to download. The announcement is on Mistral's official blog, which calls it "Le Chonk", and the same day Hugging Face published a technical guide listing every repository. The release was covered by TechCrunch and SiliconANGLE.

The model was unveiled on October 6, 2026 as an API-only preview. Over the following three weeks, Mistral says, it was improved with further reinforcement learning and customer testing, and the published weights are those of the improved version. SiliconANGLE reports that it was trained on 3,800 Nvidia Grace Blackwell GPUs in Mistral's European data centers.

CEO Arthur Mensch explained the decision by saying open source is needed so that state-of-the-art AI does not end up concentrated in the hands of a few players. Pierre Stock, VP of Science, told TechCrunch that companies and governments will be able to run it on their own premises without depending on an outside vendor that could cut off access.

Mistral Large 4 specs at a glance

Mistral Large 4 is a Mixture-of-Experts model: it has about 1 trillion parameters but uses only a small share of them for each token. Here are the figures from the official Hugging Face model card and Mistral's blog:

ItemMistral Large 4
Total parametersabout 1 trillion
Active parameters per token49 billion (52 with embeddings and output layer)
Experts128, with 4 routed per token plus 1 shared
Context1 million tokens
Input / outputtext and images / text
Languages160+, including every official EU language
Reasoning levelslow, medium, high (default)
Official formatsFP8 and NVFP4, plus GGUF for llama.cpp

The reasoning levels are handy: on low the model answers faster and uses fewer tokens, on high it thinks longer about hard tasks. The main repository is called Mistral-Large-4-1T-A52B, where "A52B" refers to the 52 billion active parameters counting embeddings and the output layer.

The license: can you use it for work?

Yes: you can use, modify, fine-tune and sell products built on Mistral Large 4, with one exception. The license is called the Mistral Ecosystem Open-Weight License, and Mistral describes it as a modified MIT license. The added condition applies to anyone offering the model to others as a service, known as Model-as-a-Service: a company selling API access to Mistral Large 4 to its own customers must first get a separate commercial license from Mistral.

In practice:

  • internal use at your company, including commercial use: allowed;
  • apps and products that embed the model: allowed;
  • fine-tuning on your own data: allowed;
  • reselling the model as an API to others: requires an agreement with Mistral.

It is more open than the LFM Open License used for Liquid AI LFM3, which has a revenue threshold, but less permissive than a plain MIT license. If your project falls under the Model-as-a-Service case, get legal advice.

What Mistral Large 4 costs and where to use it

The easiest way to use Mistral Large 4 is Mistral's API, at 0.70 dollars per million input tokens and 2.10 dollars per million output tokens. Cached input tokens cost 0.07 dollars per million. Those are the prices Mistral announced in its October 29, 2026 post.

According to the official blog, the model is available on:

  • Mistral Studio and the Mistral API;
  • OpenRouter;
  • Microsoft Foundry, with private deployment options on Azure;
  • NVIDIA, with hosted endpoints and NIM containers;
  • Hugging Face, to download the weights and run them yourself with vLLM, SGLang or llama.cpp.

Running it yourself is not a laptop job, though. The official vLLM recipe spreads the model across 8 GPUs, and the NVFP4 version requires Nvidia Blackwell cards. Hugging Face reports that a 2-bit compressed version runs on a single Mac with an Ultra chip and 512GB of unified memory at about 15 tokens per second: enough to try it, not to serve many users.

Mistral Large 4 benchmarks: what Mistral claims

The published test results come from Mistral and its launch partners and have not yet been confirmed by complete independent evaluations. The main figures claimed on the official blog:

  • 41% on Terminal-Bench 4.0, 13 points above the preview version, with 30% fewer output tokens;
  • 48% on the Artificial Analysis Coding Agent Index, using 47% fewer tokens per task: Mistral says it is the best open model for agentic coding;
  • 84.5% on BrowseComp, a web research test;
  • an average 96.6% refusal rate on harmful cybersecurity requests.

According to TechCrunch and SiliconANGLE, Mistral positions the model as close to GLM-5.3, a Chinese open model, while using fewer tokens per task. Note that Hugging Face's guide reports different numbers from Mistral's blog for some tests, likely because of different setups. Mistral says the full technical report is coming later, so treat these figures as indicative until then.

What it means for you: creators, freelancers and small businesses

For creators and small business owners, an open Mistral Large 4 mainly means more choice and lower prices, not a model to install in your office. Here is what you can actually do with it:

  1. Try it cheaply through the API. At 0.70 and 2.10 dollars per million tokens you can run plenty of tests for little money: useful for summaries, drafts and analysis of long documents thanks to the 1-million-token context.
  2. Work in your own language. Mistral claims more than 160 languages, including every official EU language, so you can compare it with Claude, ChatGPT and Gemini on your own texts.
  3. Keep data in Europe. If your clients ask for confidential data handling, a European open-weight model that a provider can install on your servers is one more option.
  4. Do not self-host it. Without a server with 8 GPUs, use the API or OpenRouter; for open models that really run on a laptop, look at Nemotron 3.5 or LFM3.

The method matters more than the model: well-written prompts get good results from any large model. The 80 prompts in Claude Business OS cover email, meetings, procedures, documents and numbers; they are written for Claude, but they make an excellent benchmark for comparing Mistral Large 4 on office tasks. If you use AI to find clients, the 70 prompts in AI Sales Machine always ask for sources, which helps with a new model. And to find your way among the tools for creating and selling digital products, there is the AI Business Toolkit.

For creative AI tools, start with our AI guides, for example the one on Gemini Omni. For a complete path there is The AI Creator System, and here is how we use artificial intelligence.

Frequently asked questions

When were the Mistral Large 4 weights released?

On October 29, 2026, on Hugging Face, in the mistralai/Mistral-Large-4-1T-A52B repository. The model had been unveiled as an API-only preview on October 6, 2026.

Is Mistral Large 4 free?

The weights are a free download and the license also allows commercial use; a separate license is needed only to offer it to others as an API service. Using it through Mistral's API costs 0.70 dollars per million input tokens and 2.10 per million output tokens.

Can I run Mistral Large 4 on my own computer?

Realistically, no: the official recipe uses 8 data-center GPUs. Hugging Face reports that a 2-bit version runs on a Mac with an Ultra chip and 512GB of memory at about 15 tokens per second. For everyday use, the API makes more sense.

Which languages does Mistral Large 4 support?

The official model card lists more than 160 languages, including every official language of the European Union.

What is "Le Chonk"?

It is the playful nickname Mistral AI gave Mistral Large 4, its largest model, with about 1 trillion parameters.

Read next

Sources checked on October 30, 2026