Mistral Large 4 open weights: what it takes to self-host

Mistral released the Mistral Large 4 open weights on October 29, 2026. The model, nicknamed "Le Chonk", has about 1 trillion total parameters with roughly 49B active per token, takes text and images, and has a 1M-token context window. It went into public preview on October 6, and the open release adds three more weeks of reinforcement learning focused on coding and defensive security.

I build agents for business software, and the question I get most from companies is "can we run this ourselves?" With this release the honest answer for a frontier-class open model is: yes, if you have serious hardware or pick a managed host you trust.

What is in the Mistral Large 4 open weights release

The announcement frames the model around two jobs: coding agents and cybersecurity defense. The headline numbers from Mistral:

Benchmark	Mistral Large 4
Artificial Analysis Coding Agent Index	48%, tied with GLM-5.3 while using 47% fewer tokens per task
Terminal-Bench 4.0	41%, up 13 points from the preview
Cognition FrontierCode 2.0	19.7%, first among the models shown
Artificial Analysis Cyber Index	51%
B3 agent security benchmark	resists 97.1% of attacks, up from 93.3% in preview

These are Mistral's own charts, and some numbers differ slightly from the Hugging Face launch post, which reports 49.8% on the Coding Agent Index and 28.3% on Terminal-Bench 4. Benchmarks depend on reasoning mode and harness. Treat them as a ranking signal and run your own evals on your own tasks.

The model now has three reasoning modes: low, medium, and high (the default). Hugging Face notes that reasoning can't be turned off. A reasoning_effort of none still runs at high.

Under the hood

From the Hugging Face post:

72 decoder layers, 128 routed experts, 4 picked per token plus one shared expert.
Multi-head Latent Attention combined with a DeepSeek Sparse Attention indexer that picks up to 2,048 earlier positions per query.
A 1.6B Pixtral vision encoder for image input.
Day-zero support in Transformers, vLLM and llama.cpp, plus fine-tuning with TRL.

Two kinds of sparsity are doing the work here. Expert routing picks which feed-forward blocks run. The attention indexer picks which context positions to look at. That is how a 1T model with a 1M context stays affordable to serve.

What it takes to self-host Mistral Large 4

This is where "open weights" meets reality. Some reference points from the Hugging Face post:

8x NVIDIA H200: the Inference Endpoints recipe runs FP8 weights with vLLM, serves up to 131K tokens of context and 16 concurrent requests, at roughly 65 tokens per second for a single request.
NVFP4 on Blackwell: vLLM supports the smaller NVFP4 checkpoint, which needs NVIDIA Blackwell GPUs.
One Mac with 512 GB of unified memory: the 2-bit GGUF runs on a single Apple Silicon Ultra. Hugging Face measured about 15 tokens per second of generation on an M3 Ultra with Q2_K.

So you can technically run it on one very large Mac, at a quality cost from 2-bit quantization. For production you are looking at a full GPU node. For most teams the realistic path is the Mistral API, OpenRouter, or a managed deployment like Microsoft Foundry, with self-hosting reserved for workloads where data must not leave your network.

The license, in plain words

The Mistral Ecosystem Open-Weight License is MIT with one condition. You can use, modify, fine-tune and commercialize the model. If your business is offering the model itself as a service, you need a commercial license from Mistral first.

For most builders that condition never applies. If you put Mistral Large 4 inside your CRM, your support tool or your internal agent, you are fine. If you plan to resell raw inference as an API, read the license before you start.

The cybersecurity angle

Mistral pitches the model hard at defenders: vulnerability reproduction, patching, malware analysis and writing detection rules. The argument for open weights here is sound. Security teams often can't send source code or incident data to an outside API, and a closed model that refuses "attack-looking" requests in the middle of an incident is a real problem.

Mistral is also open about the risk. It says it saw the model refuse black-box attacks on live targets in preview traffic, but sees potential uplift for real attacks and is updating its guardrails. A specialized offensive-security variant is only going to vetted design partners. Open weights mean anyone can remove guardrails with fine-tuning, so the vendor's refusal rate only protects you when you serve the model yourself with those guardrails in place.

My take

Two weeks ago I wrote about OpenCode 2.0 and how the harness matters as much as the model. Mistral Large 4 is the other half of that story. A strong open model with good token efficiency, plugged into an open harness, is now a real alternative to closed coding agents for teams that care about where their code goes.

For business software the interesting number is not the benchmark. It is the license and the price. When I built the AI agents plugin for erxes, the agents could only be as private as the model behind them. For a CRM holding customer data, a frontier-class model you can run inside your own network changes that conversation, at least for teams that can afford the hardware.

What I'm watching next: how fast the community produces smaller distilled versions. Mistral says Large 4 will be the base for a new generation of specialized models, and Hugging Face already shows how to fine-tune smaller models from its traces with TRL. A 30B model distilled from this for one domain might matter more to most teams than the 1T original.

FAQ

Is Mistral Large 4 free for commercial use?

Yes, under the Mistral Ecosystem Open-Weight License, with one exception: companies offering the model as a service (Model-as-a-Service) need a separate commercial license from Mistral.

How much does the Mistral Large 4 API cost?

$0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens, according to Mistral.

Can I run Mistral Large 4 locally?

Only on big hardware. A 2-bit GGUF runs on a single Apple Silicon Ultra with 512 GB of memory at around 15 tokens per second. Production setups use a node of 8 H200s or Blackwell GPUs.

Sources

Mistral Large 4 (Le Chonk) is now open weight, Mistral AI, October 29, 2026
Welcome Le Chonk: Mistral Large 4 on Hugging Face, Hugging Face, October 29, 2026
mistralai/Mistral-Large-4-1T-A52B model card on Hugging Face

Cover image: Mistral AI