Mistral Large 4 open weights: Le Chonk is a 1-trillion-parameter model anyone can download

Mistral Large 4 is now open weight. On 29 October, Mistral AI published the full weights of its largest model, nicknamed "Le Chonk", on Hugging Face, three weeks after a public preview. It is a 1-trillion-parameter model that anyone can download, fine-tune and run, and Mistral is pitching it at two jobs: writing code and defending networks.

Image: Mistral AI (announcement page)

What is Mistral Large 4?

Mistral Large 4 is a general model that handles chat, reasoning, coding and agent work in one set of weights. According to the model card, it reads text and images, writes text, works in more than 160 languages and has a 1-million-token context window. You can pick how hard it thinks per request: low, medium or high (the default).

Mistral says it trained the model from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own data centres in Europe, and that the open release got about three more weeks of reinforcement learning after the preview.

How does a 1-trillion-parameter model run?

The trick is a design called mixture of experts (MoE). Instead of one giant network that fires in full for every word, the model holds 128 smaller "expert" networks plus one shared expert. For each token, a router picks 4 of the 128. So while the download holds about 1 trillion parameters, only about 49 billion do work per token (52 billion counting the embedding and output layers), roughly 5% of the model.

Chart: NextDiff, from Mistral AI's model card

That keeps each answer cheaper to compute than a dense model of the same size. It does not shrink memory needs: all the weights still have to sit on the GPUs. Mistral also ships a compressed NVFP4 checkpoint for Nvidia's Blackwell chips, which The Next Web reports runs on four B200 or B300 GPUs. That is server hardware, not a gaming PC.

How good is Mistral Large 4 at coding?

On Mistral's own charts, it is at or near the top of the open models it compared:

• Terminal-Bench 4.0 (agentic coding in a terminal): 41%, up from 28% for the preview, just ahead of GLM-5.3 at 40% and DeepSeek V4.1 Flash at 31%.
• Artificial Analysis Coding Agent Index: 48%, "effectively tied" with GLM-5.3, while using 47% fewer tokens per task.
• Cognition's FrontierCode 2.0: 19.7%, the top score in Mistral's chart.
• Blind human preference (Surge): Mistral says only Claude Opus 5.5 ranked higher among the models tested.

Image: Mistral AI (vendor-reported benchmark)

Mistral also showed the model building apps on its own. In one demo it wrote "Orbital Trace", a live 3D tracker of satellites and space debris from public orbit data, then tested and sped up its own code.

Image: Mistral AI (agentic coding demo)

The caveat: on the broader Artificial Analysis Intelligence Index v5, Mistral says Large 4 scores 26, the highest of any Western open-weight model, but admits that the leading Chinese open models score higher overall. As with Kimi K3.1, these are mostly vendor-reported numbers.

Why is Mistral pitching it at cybersecurity?

Mistral says the model can find a bug in a codebase, reproduce it and write a patch without breaking anything, and that it refused 96.6% of clearly malicious cyber requests in its tests, the highest rate it measured among open models. It scores 51% on Artificial Analysis's Cyber Index, though two models in Mistral's own chart score higher there.

A more capable variant for penetration testing and red-teaming is only in a vetted preview, restricted to systems the tester owns or is allowed to test. The open weights carry no such gate. As Mistral’s Pierre Stock told TechCrunch, access to an open model "can't be cut off by definition".

How much does Mistral Large 4 cost?

The weights are free under the Mistral Ecosystem Open-Weight licence, a modified MIT licence. You can use, modify, fine-tune and sell products built on it; only companies running it as a hosted model-as-a-service need a separate commercial licence.

On Mistral's API it costs $0.70 per million input tokens ($0.07 cached) and $2.10 per million output tokens. It is also on OpenRouter, Nvidia's NIM endpoints and Microsoft Foundry, and runs with vLLM and SGLang.

Why it matters

The open models at the top of the leaderboards mostly come from Chinese labs such as Zhipu (GLM), Moonshot (Kimi) and DeepSeek, as Mistral’s own charts show. Mistral now offers a European model at the trillion-parameter scale, with a licence that lets banks, governments and security teams run it on their own hardware and keep their data in-house. That "sovereign AI" pitch is why customers such as ABN AMRO and CrowdStrike are quoted in the launch.

What's next

Mistral says a full technical report is coming. Independent tests will show whether the coding lead over GLM-5.3, one point on Terminal-Bench, holds up outside Mistral's own charts.