Mistral Large 4 open weights: a 1T-parameter MoE model you can download

On October 29, Mistral released the open weights of Mistral Large 4 (also nicknamed “Le Chonk”), its flagship model with about 1 trillion parameters. The Mistral Large 4 open weights are already on Hugging Face, and the model is also available through Mistral’s API and OpenRouter.

What’s new about the Mistral Large 4 open weights?

Mistral Large 4 showed up in early October as a public preview, but back then only through the API. Now we get a version with additional training and open weights. According to the model card, it is:

1. An MoE (mixture-of-experts) model with about 1T parameters, of which about 52B are active per token (128 experts, 4 active plus 1 shared).
2. A model with a 1M-token context window.
3. A multimodal model: it takes text and images as input and answers in text.
4. A hybrid model with three reasoning levels: low, medium and high (the default).

According to the announcement, the model is fluent in more than 160 languages. Interestingly, Mistral trained it on 3,800 NVIDIA Grace Blackwell GPUs in Europe.

What can Mistral Large 4 do?

Mistral is betting mainly on coding, agents and cybersecurity. Results from the announcement:

Benchmark	Score
Terminal-Bench 4.0	41% (preview: 28%)
Artificial Analysis Coding Agent Index	48%
FrontierCode 2.0	19.7%
Artificial Analysis Cyber Index	51%
BrowseComp	84.5%
B3 Agent Security Benchmark	97.1% (preview: 93.3%)

The most interesting part is the jump on Terminal-Bench: 13 percentage points in three weeks, while using 30% fewer output tokens (according to Mistral). These are the company’s own numbers, so let’s take them with a grain of salt until we see independent tests.

How much does it cost, and what’s the license?

API pricing from the announcement:

1. $0.70 per 1M input tokens,
2. $0.07 per 1M cached input tokens,
3. $2.10 per 1M output tokens.

The weights are released under the Mistral Ecosystem Open-Weight license, a modified MIT license. We can use, modify, fine-tune and commercially use the model, but a company that wants to sell it as a service (Model-as-a-Service) needs a separate commercial license from Mistral.

Can we run Mistral Large 4 locally?

In theory, yes. In practice, probably not at home. The examples in the model card assume 8 GPUs, and according to The Next Web, the quantized version runs on four NVIDIA B200 or B300 cards. Mistral has also prepared an NVFP4 checkpoint for Blackwell GPUs, and the model is supported by vLLM, SGLang, Transformers and llama.cpp (GGUF builds are already out).

So running a trillion-parameter model on a laptop is more of a curiosity than a real scenario ;) If we want to play with models locally, it’s better to start with smaller ones, which I cover in my guide to running AI locally.

How to try Mistral Large 4?

We have a few options:

1. Vibe Code at chat.mistral.ai.
2. The API through Mistral Studio (console.mistral.ai) or OpenRouter.
3. The weights from Hugging Face, if we have access to a suitable GPU cluster.
4. NVIDIA endpoints and the NIM container, plus Microsoft Foundry.

It’s another big open-weight model in recent weeks, after Reflection Beam and the announcement of Kimi K3.1 Pro. Earlier in October, Mistral also showed Voxtral 3, which, in turn, didn’t get open weights.

FAQ

Is Mistral Large 4 open source?

It has open weights under the Mistral Ecosystem Open-Weight license (a modified MIT license). Commercial use is allowed, except for offering the model as a service without a separate license from Mistral.

How many parameters does Mistral Large 4 have?

About 1 trillion parameters in an MoE architecture, of which about 52B are active per token (49B without the embedding and output layers).

Does Mistral Large 4 understand images?

Yes, it takes text and images as input and answers in text.

How much does the Mistral Large 4 API cost?

$0.70 per 1M input tokens and $2.10 per 1M output tokens (according to the pricing in the announcement).