Mistral Large 4

What it is

Mistral Large 4 is the flagship model of Mistral AI. It is a large language model that is natively multimodal, meaning it was built to handle more than one kind of input, such as text and images, rather than adding image handling afterwards^[1][2]. Its nickname is "Le Chonk"^[2]. The model has about 1 trillion total parameters and 52 billion active parameters, and its Hugging Face model ID is Mistral-Large-4-1T-A52B^[1][2]. Total parameters are all the learned values in the model; active parameters are the ones used to process any single request.

Mistral's changelog describes Mistral Large 4 as a multimodal mixture-of-experts model with a 1M-token context window^[3]. A mixture-of-experts design splits the network into many specialist parts and routes each input to only some of them. A token is a small unit of text, roughly a fragment of a word. A context window is the amount of text the model can take into account in a single request.

Origins

Mistral AI announced the model as a public preview on 6 October 2026^[1]. On 29 October 2026 it released the weights of an improved version under the Mistral Ecosystem Open-Weight license^[1][2][3]. Open weights means the trained parameters are published so that others can download and run the model. The release replaces Mistral Large 3 and Medium 3.5^[2][4].

Mistral trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its European data centres^[1]. Training from scratch means starting with random values rather than building on an earlier model.

Main features

The model has three selectable reasoning modes: Low, Medium and High, with High as the default^[2]. A reasoning mode controls how much work the model spends thinking before it answers.

Its licence is a modified MIT licence. Under it, a separate licence is required only for offering the model as a service to others^[2]. The MIT licence is a short, permissive open-source licence; the modified version adds the service restriction described here.

Mistral emphasised coding and cybersecurity^[2]. It reported a 96.6% average refusal rate on malicious cyber requests^[2]. A refusal rate is the share of harmful requests the model declines to fulfil.

Mistral's reported scores for the open-weight release were 41% on Terminal-Bench 4.0, which is 13 points above the preview; 48% on the Artificial Analysis Coding Agent Index; 84.5% on BrowseComp; and 97.1% on B3 agent-security attack resistance^[2]. These are the company's own figures. Benchmarks are standard test sets, and a single score does not capture every kind of use.

The model is available on Mistral Studio, the Mistral API, Hugging Face, OpenRouter, NVIDIA NIM and Microsoft Foundry^[2][4]. Its API price is $0.70 per million input tokens, with $0.07 per million for cached input, and $2.10 per million output tokens^[1].

Place in development

Mistral Large 4 sits at the top of Mistral's line of models. It is positioned for agentic coding, meaning coding work in which the model plans and runs multi-step tasks with tools, and for cybersecurity^[1][2].

Readers should note that benchmark scores and prices come from Mistral and can change. The Mistral changelog and model overview are the places where later updates should be checked.