What hardware do I need to run Mistral Large 4?

At its most compressed (IQ1_S (1-bit)) it needs roughly 233GB of VRAM. VRAM figures are approximate and depend on context length and settings.

Can I run Mistral Large 4 on Apple Silicon?

It can run on Apple Silicon through general runtimes, but it is not specifically optimised for it.

What is Mistral Large 4's context window?

Mistral Large 4 has a context window of 1,000,000 tokens, about 1000k.

Mistral Large 4

Name: Mistral Large 4
Author: Mistral AI

Mistral AI · Vision-language · MoE ~1T (A49B) · 1000k context · Released 29 October 2026

Permitted with conditions Open weights

Mistral Large 4 launched as an API preview on 6 October 2026 and had its open weights published on 29 October. It is a hybrid instruct-and-reasoning mixture-of-experts of around one trillion total parameters, with 72 decoder layers (the first dense, the rest MoE), a router selecting four of 128 experts plus one always-on shared expert, and roughly 49 billion active parameters per token (52 billion including embeddings, hence the "A52B" in the repository name). It accepts text and images, using a 1.6-billion-parameter Pixtral vision encoder, and produces text, with a context window up to one million tokens (deployment examples commonly cap at 131k). Mistral ships FP8 and NVFP4 checkpoints, with community GGUF builds from the llama.cpp team and Unsloth, and points to vLLM, SGLang, an NVIDIA NIM container and TensorRT-LLM for serving. At this scale it is a large-memory model: a 2-bit quant fits a single 512GB Apple Silicon machine, while higher-quality quants need distributed inference.

Strengths

A trillion-parameter multimodal open-weight model from a major European lab, self-hostable with data kept in-house
One of the most permissive recent conditional licences, with Model-as-a-Service the only real gate
Very long one-million-token context window
Sparse mixture-of-experts, so only about 49B of the ~1T parameters compute per token
Official FP8 and NVFP4 checkpoints plus community GGUFs down to 1- and 2-bit

Weaknesses

Large-memory footprint, needing a 512GB-class machine even at 2-bit and distributed inference for good-quality quants
A 2-bit quant is the only single-machine route, and 1- to 2-bit is a real quality compromise for a frontier model
Capability figures at launch are largely Mistral's own, with independent evaluation still limited
Multimodal input only, text output

Hardware requirements

Quantisation	Approx. VRAM	Notes
IQ1_S (1-bit)	~233GB	The smallest community GGUF, around 233GB, fits a 256GB machine, but 1-bit is a steep quality compromise for a frontier model. Figure is approximate, from Unsloth's GGUF repo.
Q2_K_XL (2-bit)	~378GB	A 2-bit quant, around 378GB, is the single-machine route: it fits a 512GB Apple Silicon Ultra. Still a real quality trade for reasoning. Figure is approximate.
Q4_K_XL (4-bit)	~605GB	A 4-bit quant, around 605GB, is the sensible-quality point, needing roughly 768GB or more, or multi-GPU distributed inference. Figure is approximate.
Q8_0 / FP8 (8-bit)	~1120GB	Near full precision, around 1.12TB, is multi-node server territory. The official FP8 checkpoint is in this class. Figure is approximate.
BF16	~2100GB	Full precision, around 2.1TB, for setups that need it. Figures are approximate and rise with context and concurrency.

What you'd need to run this

Licence

Mistral Ecosystem Open-Weight License — read the licence

What this means for commercial use

Open weights under the new Mistral Ecosystem Open-Weight License, which Mistral describes as a modified MIT licence. Read at source, it grants broad rights to use, modify, fine-tune, distribute, sublicense and sell, and the only condition beyond retaining the MIT-style copyright notice is on Model-as-a-Service: a business that operates a Model-as-a-Service offering (giving third parties inference or post-training access with meaningful control over inputs, parameters or training data) must obtain a separate commercial licence from Mistral before any commercial use. Internal use, and embedding the model inside an end-user product where its capabilities sit behind specific features, are explicitly exempt. There is no revenue threshold and no territory clause. For almost everyone except model-hosting providers this is effectively permissive. We are describing the licence, not giving legal advice; read it before relying on it.

Mistral Large 4: common questions

What hardware do I need to run Mistral Large 4?: At its most compressed (IQ1_S (1-bit)) it needs roughly 233GB of VRAM. VRAM figures are approximate and depend on context length and settings.
Is Mistral Large 4 free for commercial use?: Commercial use is permitted, but with conditions. Open weights under the new Mistral Ecosystem Open-Weight License, which Mistral describes as a modified MIT licence. Read the licence before relying on it at scale.
Can I run Mistral Large 4 on Apple Silicon?: It can run on Apple Silicon through general runtimes, but it is not specifically optimised for it.
What is Mistral Large 4's context window?: Mistral Large 4 has a context window of 1,000,000 tokens, about 1000k.

Availability

Recommended for

Organisations with a 512GB-class machine or multi-GPU infrastructure wanting a frontier-scale open multimodal model
Self-hosting where data must stay in-house and the only likely licence question is whether you run a Model-as-a-Service
Long-context and long-document multimodal work up to one million tokens

Related models

Run it with

Related guides

Glossary

Our coverage

Mistral Large 4 'Le Chonk' open weights have landed, under a new modified-MIT licence

Catalogue entry last verified 1 November 2026. Specifications change; verify anything you are about to spend money on.