Mistral kept its promise. On October 29, three weeks after the API-only preview, the Paris lab published the full open weights of Mistral Large 4, the trillion-parameter model it calls Le Chonk. No waitlist, no gate, no provider standing between you and the model. You can pull it from Hugging Face and run it on your own hardware today.

The scale numbers are the headline. It carries roughly 1 trillion total parameters (1.05 trillion in Mistral's docs) with only 49 billion active per token in a natively multimodal mixture-of-experts design, and a 1-million-token context window. Mistral trained it from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European datacenters. The API got cheaper the same day: about $0.70 per million input tokens and $2.10 per million output tokens, roughly half the preview price.

What's new since the preview

This isn't the October 6 checkpoint repackaged. Mistral says three more weeks of reinforcement learning, plus testing with governments, corporate users and security partners, improved it substantially. Terminal-Bench 4.0 is up to 41% from 28.3% at preview. Mistral also claims 48% on Artificial Analysis's Coding Agent Index, effectively tied with China's GLM-5.3, while using 47% fewer tokens per task. As always with vendor numbers, treat them as a starting point, not a verdict.

The independent read is kind. Artificial Analysis gives Large 4 26 points on its Intelligence Index v5, which makes it the highest-scoring open-weight model developed outside China, one point above Meta's Muse Spark Open. It also reports the cost per completed task at $2.06, down from $5.54 for the preview. That is a 63% cut, and it's the part competitors should worry about most.

The catch in the license

Here's the catch. Large 4 ships under Mistral's Ecosystem Open-Weight License, a modified MIT license with one important carve-out. You can use it, fine-tune it, and sell products built on it. But any business offering the model itself to third parties as a hosted service must buy a separate commercial license from Mistral. And one practical note: free weights don't mean free hosting. A 1-trillion-parameter model still needs serious hardware, so most people will still go through Mistral Studio, OpenRouter, Microsoft Foundry or an Nvidia NIM container.

Mistral is pitching this as a cyber-defender's model, and the numbers are interesting. Its own blog reports 82% on Artificial Analysis's test that asks a model to reproduce a real vulnerability and then patch it, the highest score of any model, plus 93% of Cybench's security challenges. Its average refusal rate on malicious cyber requests sits at 96.6%. The company is honest about the trade-off it can't engineer away: open weights can't be recalled, so a stronger pen-testing version stays restricted to vetted partners.

Our take

The contrast this week is hard to miss. Google announced Gemini 4 Carbon on October 28 and is keeping it for vetted cyber defenders and partners for now. Mistral took the opposite bet and published everything, weights included. As Mistral's VP of science Pierre Stock told Reuters: "I'm not comfortable living in a world in which open weights do not exist." We checked the key facts against Mistral's own blog and pricing pages, and every figure above is labeled for what it is, vendor claim or independent test. With Kimi K3.1 Pro's weights promised by November 4, the open-versus-closed race is about to have its loudest week of the year. Builders win either way.