Mistral opens Large 4 weights: a 1T model you can own

Open weights just caught up with the work you actually do.

Mistral released the weights of Mistral Large 4, the model it nicknames Le Chonk, on Oct. 29. It has about a trillion parameters, and Mistral calls it the best open-weight model for coding. You can download it, run it on your own hardware and keep running it whatever happens to anyone's API.

A trillion parameters, 49 billion at a time

The Hugging Face model card lists about 1 trillion total parameters, with 49 billion active per token (52 billion counting embeddings and output layers). That is a mixture of experts design: only a slice of the model works on each token. The card lists a context window of 1 million tokens.

Mistral reports these scores for the final release:

• Terminal-Bench 4.0: 41%, up from 28% for the preview.
• Artificial Analysis Coding Agent Index: 48%, which Mistral says is effectively tied with China's GLM-5.3 while using 47% fewer tokens per task.
• Artificial Analysis Cyber Index: 51%.

The closed models are still ahead. Fortune reports that Anthropic's Mythos 5.1 scored 60.9% on the same Terminal-Bench test.

When TopFive covered the preview on Oct. 6, the question was whether Mistral could close the gap with China's open models. Fortune says three more weeks of reinforcement learning took the Terminal-Bench score from 28% to 41%.

Thirteen points in three weeks.

Owning the file is the point

Yes, it trails the best closed models. That misses why this matters.

Weights you hold can't be switched off. Pierre Stock, Mistral's VP of science, made that case to Fortune, which points to a June US export-control directive that forced Anthropic to suspend access to two of its models for nearly three weeks. Once weights are public, Stock told Fortune, "it's impossible to cut off access."

The license is the other half. Mistral ships Large 4 under its Mistral Ecosystem Open-Weight License, which the model card describes as a modified version of the MIT license, for commercial and non-commercial use. Selling it as a model-as-a-service business needs a separate commercial license from Mistral.

If you have ever had a vendor change terms on you, you know which half matters more.

What it costs to run, and where

You don't have to host it. Mistral lists API pricing of $0.70 per million input tokens, $0.07 for cached input and $2.10 for output.

$2.10 per million output tokens.

Mistral says Large 4 is available in Mistral Studio, on its API, on OpenRouter, on NVIDIA-hosted endpoints and in Microsoft Foundry.

Self-hosting is a data center job, not a laptop job. The model card's serving examples for vLLM and SGLang assume a node of eight GPUs, and an NVFP4 checkpoint targets Nvidia's Blackwell chips. The page also lists GGUF and other quantized builds for llama.cpp, Ollama and LM Studio.

The cyber question is still open

Mistral is pitching Large 4 to defenders, and it is candid about the risk. The company reports a 96.6% average refusal rate on malicious cyber requests, which it calls the highest among the open models it measured. It also says that after the preview it saw "a potential for uplift to conduct real attacks," and that it is changing its guardrails. A specialized cyber variant stays in preview for vetted partners such as Orange Cyberdefense.

Open weights cut both ways here. The guardrails ship inside the file, and whoever holds the file decides what to keep.

Watch the Artificial Analysis leaderboard as independent runs come in. If the token savings hold, the price per finished task drops further than the list price suggests.

Pull it into your evals this week. Then price it against your current bill.