What Is Mistral Large 4? Le Chonk’s Open Weights, Price and License

Mistral Large 4 is a 1-trillion-parameter AI model from Mistral AI, and its weights are now free to download. Mistral opened a public preview on its API on October 6, 2026, and announced the open weights on October 29.

Mistral calls the model Le Chonk. It reads text and images and holds a 1 million token context window. On Mistral’s API it lists at $0.70 per million input tokens and $2.10 per million output tokens. This post covers what shipped, the two release dates, the license and the price. It also covers where you can use it and what it means for agencies and nonprofits picking a model.

The short version: Mistral Large 4 is cheap to call and free to run yourself for internal use. Mistral says it was trained and is served in Europe. The catch is in the license. If you sell access to the model as a service, you need a separate commercial license from Mistral. The API is also still labeled public preview as of November 2026.

What Mistral shipped with Mistral Large 4

Mistral Large 4 is a mixture-of-experts (MoE) model. That means only a small part of the network runs for each token. The Hugging Face model card lists 128 routed experts, with 4 active per token plus 1 shared expert. It puts the active count at 49 billion parameters per token, or 52 billion counting embeddings and output layers. That second number is where the repository name, A52B, comes from.

The Mistral Large 4 docs page also lists structured outputs, function calling, document Q&A, batching and agents among its capabilities. Mistral’s announcement says the model supports more than 160 languages.

Our explainer on what a context window is covers what 1 million tokens holds and what it costs.

Two dates: the October 6 preview and the October 29 weights

Mistral released this model in two steps. Mistral’s API changelog says Mistral Large 4 became available in public preview on October 6. That entry also said "Open weights are coming soon" and offered launch pricing at 50% off for two weeks.

Mistral’s October 6 announcement promised the weights by the end of the month. On October 29, Mistral published a second post announcing improved weights. That post refers to the preview as released "earlier this month." Mistral reports the new weights score higher on several agent and coding tests than the preview did.

On Hugging Face, the files arrived over the following days. As of November 2, both the main checkpoint and a smaller NVFP4 checkpoint were posted. The main repository is about 1.06 TB. The NVFP4 version, which targets NVIDIA Blackwell GPUs, is about 610 GB.

One detail matters for planning. The docs page still lists the API model as "Public Preview," version v26.10, as of November 2026. Mistral has not posted a general availability date.

The Mistral Large 4 license: modified MIT, with one condition

Mistral released the weights under the Mistral Ecosystem Open-Weight License. Mistral’s announcement calls it "a modified MIT license." Most of it reads like MIT. You may use, copy, modify, distribute and sell the model, as long as you keep the copyright notice.

The change is a single condition. The license file says that if a licensee or its affiliates operate "a Model as a Service business," then "the Licensee must enter into a separate commercial license with Mistral before using the Model or its derivative works for any commercial purpose."

The license defines Model as a Service as "giving a third party access to the Model or its derivative works for inference or post-training by any means (for instance via API or hosted endpoints), in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data."

Two things fall outside that definition, per the license text: "end-user products with model capabilities solely embedded within specific features or harnesses," and "mere relaying of requests to models hosted by others." The condition also does not apply to "internal use of the Model or its derivative works" or to use through Mistral’s official products.

For most of our readers, internal use and features built into a product look clear. Reselling raw model access looks like the case that needs a talk with Mistral. The license points questions to Mistral’s contact form. Our guide to open vs closed AI models explains why "open weights" licenses vary this much.

What Mistral Large 4 costs and where to use it

Mistral’s docs list the API price in US dollars, with euros as an option:

Those are list prices as of November 2026. Mistral’s October 6 changelog described a two-week launch discount, which has ended.

Mistral’s October 29 post names these places to use the model:

• Mistral Studio and the Mistral API.
• OpenRouter.
• NVIDIA-hosted endpoints, plus an NVIDIA NIM container.
• Microsoft Foundry.
• Vibe Code, which Mistral links on chat.mistral.ai.
• Hugging Face, for the open weights.

Mistral’s posts do not say whether the regular Le Chat assistant uses Mistral Large 4. They do not name AWS or Google Cloud either. If you need one of those, check that platform’s own model list.

Our breakdown of what AI actually costs shows what a task costs at rates like these.

Data residency: what Mistral says about Europe

Mistral puts Europe at the center of this release. The October 6 post says the preview runs on Mistral’s own European datacenters. It also says the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in those datacenters.

That post describes a planned European deployment that Mistral operates end to end, "independently of other digital service providers and under European law." It is a plan, and the post gives no date.

Mistral’s help center covers where API data sits today. Its article on where API data is stored says: "Mistral AI’s servers are located in Europe (Sweden)."

Two limits apply. That help article is not specific to Mistral Large 4. And if you call the model through OpenRouter, NVIDIA or Microsoft Foundry, that provider’s hosting terms apply instead of Mistral’s.

The open weights give a second route. If your data cannot leave your own systems, you can run the model on hardware you control. That is a large server job, not a laptop job. The main checkpoint is about 1.06 TB, and Mistral’s serving examples use a node of eight GPUs.

What Mistral claims on benchmarks

Every score below is Mistral’s own, from its October 29 post. No independent lab had published a full test of the final weights when we checked.

Mistral reports 41% on Terminal-Bench 4.0, up 13 points from the preview. It reports 48% on the Artificial Analysis Coding Agent Index, which it calls roughly tied with GLM-5.3. It says Mistral Large 4 ranked behind only Claude Opus 5.5 on Surge’s human preference test for code.

Mistral also leans on security. It reports that the model refused 96.6% of malicious cyber requests on average. Mistral titled the post "Built for Coders, Trusted by Defenders." A full technical report is listed as coming soon. Treat these numbers as a starting point for your own tests.

What Mistral Large 4 means for agencies and nonprofits

For a small nonprofit, the first fact is price. At $2.10 per million output tokens, Mistral Large 4 is cheap to test. For drafting, summaries and staff questions, a test month will tell you more than any benchmark.

For organizations with members, donors or staff in Europe, the second fact is location. Mistral says its servers are in Sweden and that the model was trained in Europe. If your data policy prefers EU processing, check that fact against Mistral’s data agreement.

For agencies, the license is the main question. Building Mistral Large 4 into a feature of a client site or app looks like the "embedded" case. Running your own endpoint and selling access to it looks like Model as a Service. Ask Mistral before you build a business on the second.

The open weights also act as an exit plan. If API prices change, the same model is on OpenRouter, NVIDIA and Microsoft Foundry, or on your own servers. That is a real hedge for a nonprofit that plans in budget years.

Two cautions apply. The API is still in public preview. Keep critical work on a stable model until Mistral announces general availability. If you only need a small model on your own machine, see our guide to local AI models. Mistral’s smaller releases, like Mistral OCR 4, may also fit a narrow task better.

How Mistral Large 4 compares with Kimi K3 and Qwen3.8

Mistral Large 4 joins a short list of open models at the trillion-parameter scale. Moonshot AI’s Kimi K3, covered in our Kimi K3 explainer, has 2.8 trillion parameters. Moonshot listed it at $15 per million output tokens when we wrote about it. Alibaba previewed Qwen3.8 as a 2.4-trillion-parameter model and promised open weights, as we covered in What Is Qwen3.8?. Those weights had not shipped as of our August follow-up. Both of those labs are in China. Mistral Large 4 is the smallest of the three by total size. It is also the only one from a European lab.

Frequently Asked Questions