Mistral AI unveils Mistral Large 4, a trillion-parameter open-weight model

Background

Mistral AI develops large language models and sells access to them through its own API and partner platforms. Mistral Large 4 is the company's flagship model in the autumn of 2026, and this entry covers the two announcements that introduced it.

What happened

On 6 October 2026 Mistral AI opened a public preview of Mistral Large 4^[1]. The model has about 1 trillion total and 52 billion active parameters, and it is natively multimodal, meaning it was built to handle several kinds of input, such as text and images^[1][2].

On 29 October 2026 Mistral released the weights of an improved version, nicknamed "Le Chonk", under the Mistral Ecosystem Open-Weight license^[3][2]. Releasing weights means that the trained model parameters are published for others to download and run.

Key details

The model was trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European data centres^[1]. A GPU is a processor widely used to train large AI models.

API pricing is $0.70 per million input tokens and $2.10 per million output tokens^[1]. An API price is the cost charged per unit of text processed, usually counted in millions of tokens.

Mistral positioned the model for agentic coding and cybersecurity^[2]. Agentic coding refers to coding work in which the model plans and carries out multi-step tasks, often with tools.

The model is available through Mistral's own platform and through outside providers. The sources for this entry are Mistral's own news pages, which are the primary sources for the announcements^[1][2][3].

Why it matters

The two announcements show a large model moving from a closed preview to an open-weight release within about three weeks. For developers and researchers, the open weights make it possible to inspect and run a model of this scale, if they have the hardware. Most users will still reach it through an API or a hosted platform.

Limits

The figures in this entry come from Mistral's own announcements. They describe the model's size, training hardware, pricing and intended uses. Independent measurements, and any later changes to price or licence terms, should be checked against later sources.