Can I use Mistral Large 4 weights commercially?

Mistral Large 4 ships under the Mistral Ecosystem Open-Weight license, a modified MIT license. Mistral says anyone can use, modify, fine-tune and commercialize the model for free. The one exception is a business that runs a Model-as-a-Service offering: it needs a separate commercial license from Mistral.

What hardware does Mistral Large 4 need to run locally?

Mistral Large 4 is a roughly 1T-parameter model, so the official FP8 and NVFP4 checkpoints are served across 8 GPUs with vLLM or SGLang, and the NVFP4 build needs NVIDIA Blackwell GPUs. Hugging Face reports that a 2-bit GGUF build fits on a single Apple Silicon Ultra with 512 GB of unified memory, at about 15 tokens per second.

How is the open-weight Mistral Large 4 different from the October preview?

The open-weight Mistral Large 4 had three more weeks of reinforcement learning than the public preview from early October. Mistral reports Terminal-Bench 4.0 rising from 28% to 41% with 30% fewer output tokens, Finch rising from 67.4% to 73.3%, and attack resistance on the B3 agent-security test rising from 93.3% to 97.1%.

How does Mistral Large 4 compare with other open-weight models?

Mistral says Mistral Large 4 scores 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3, while using 47% fewer tokens per task. Its model card gives an Artificial Analysis Intelligence Index v5 score of 26, the highest among western open-weight models, but leading Chinese open-weight models score higher on that aggregate.

Mistral AI · 2026-10-29 · major

Mistral Large 4 open weights — Le Chonk lands on Hugging Face

Mistral Large 4 (Le Chonk) is now open-weight: Mistral released the 1T-parameter multimodal MoE on Hugging Face after three more weeks of RL. It scores 41% on Terminal-Bench 4.0, up 13 points from the preview.

Cover art for Mistral's Le Chonk open-weight release of Mistral Large 4

Mistral's trillion-parameter flagship is now downloadable, with better agent and coding scores than the preview.

Key specs

Terminal bench 4.0	41%
Aa coding agent index	48%

Quick facts

Maker	Mistral AI
Weights	mistralai/Mistral-Large-4-1T-A52B (FP8, NVFP4)
Parameters	~1T total · 49B active (MoE)
Context window	1M tokens
License	Mistral Ecosystem Open-Weight (modified MIT)
API price	$0.70 in / $2.10 out per 1M tokens
Reasoning modes	Low, Medium, High (default)

What is it?

Mistral AI released the weights of Mistral Large 4, nicknamed Le Chonk, on October 29, 2026, three weeks after the API-only public preview. The released model is an improved version: it had three more weeks of reinforcement learning, and the old "None" reasoning mode is replaced by Low, Medium and High.

How does it work?

Under the hood is a sparse mixture-of-experts with about 1T parameters, of which roughly 49B run per token: a router picks 4 of 128 experts plus one shared expert. It reads text and images through a 1.6B Pixtral vision encoder and handles up to 1M tokens. Official FP8 and NVFP4 checkpoints are on Hugging Face, and the model card gives ready vLLM and SGLang recipes for 8-GPU servers.

Why does it matter?

Security teams and regulated companies can now run a frontier-class model on their own hardware instead of sending data to a US or Chinese API. Mistral reports a 96.6% refusal rate on malicious cyber requests and has partners such as ENISA, Cisco and CrowdStrike testing it, while a stronger cyber variant stays limited to vetted design partners.

Who is it for?

self-hosting teams, security teams, European enterprises

Frequently asked questions

Can I use Mistral Large 4 weights commercially?: Mistral Large 4 ships under the Mistral Ecosystem Open-Weight license, a modified MIT license. Mistral says anyone can use, modify, fine-tune and commercialize the model for free. The one exception is a business that runs a Model-as-a-Service offering: it needs a separate commercial license from Mistral.
What hardware does Mistral Large 4 need to run locally?: Mistral Large 4 is a roughly 1T-parameter model, so the official FP8 and NVFP4 checkpoints are served across 8 GPUs with vLLM or SGLang, and the NVFP4 build needs NVIDIA Blackwell GPUs. Hugging Face reports that a 2-bit GGUF build fits on a single Apple Silicon Ultra with 512 GB of unified memory, at about 15 tokens per second.
How is the open-weight Mistral Large 4 different from the October preview?: The open-weight Mistral Large 4 had three more weeks of reinforcement learning than the public preview from early October. Mistral reports Terminal-Bench 4.0 rising from 28% to 41% with 30% fewer output tokens, Finch rising from 67.4% to 73.3%, and attack resistance on the B3 agent-security test rising from 93.3% to 97.1%.
How does Mistral Large 4 compare with other open-weight models?: Mistral says Mistral Large 4 scores 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3, while using 47% fewer tokens per task. Its model card gives an Artificial Analysis Intelligence Index v5 score of 26, the highest among western open-weight models, but leading Chinese open-weight models score higher on that aggregate.

Try it

vllm serve mistralai/Mistral-Large-4-1T-A52B (see the model card for the 8-GPU flags)