Mistral Releases Open Weights for Its Trillion-Parameter Large 4 Model

Mistral AI has published downloadable weights for Mistral Large 4, nicknamed Le Chonk, turning its trillion-parameter flagship from an API-only preview into a model organizations can run on their own infrastructure, with a license that still reserves hosted model services for a separate commercial deal.

Mistral AI has released the downloadable weights of Mistral Large 4, the trillion-parameter flagship it first showed as a public preview on October 6. The release, published on October 29, lets organizations download the model from Hugging Face and run it on their own servers, instead of only calling it through Mistral’s API. The French company nicknames the model Le Chonk, and it is pitching the release as a milestone for open-weight AI in Europe: a frontier-scale model that governments and companies can deploy inside their own firewalls, with no vendor in a position to switch it off remotely.

The distinction matters because the October 6 preview was API-only. Developers could try Large 4 through Mistral Studio and the Mistral API, but they could not download, inspect, or modify the model itself. Open weights change that relationship. The weights are the numerical settings learned during training, and having them means an organization can host the model, fine-tune it for a specific job, and keep its data on infrastructure it controls. That is the point of the release, and it is also the reason the release comes with conditions that are easy to miss in the headline.

Large 4 is a mixture-of-experts model with about 1.05 trillion total parameters, of which roughly 52 billion are active for any given token when embedding and output layers are counted, or about 49 billion active per token on the narrower count. It accepts text and image input and carries a 1 million-token context window, putting it in the same size class as the largest open-weight models from Chinese labs. Mistral says the released version is an improved model, not simply the preview with a download link attached: it credits three additional weeks of reinforcement learning after the preview for gains in coding, defensive cybersecurity, and reasoning efficiency.

The numbers Mistral reports are its own, and they should be read that way. On the Artificial Analysis Coding Agent Index, Mistral says Large 4 scores 48 percent, effectively tied with China’s GLM-5.3 as the top open model on that index, while using 47 percent fewer tokens per task on average. On Terminal-Bench 4.0, a benchmark of terminal-based coding tasks, Mistral reports a rise from 28 percent in the preview to 41 percent in the released model, a 13-point gain achieved while cutting output tokens by 30 percent. A full technical report is still, in Mistral’s words, coming soon, so independent researchers have not yet had the chance to reproduce these results on the released weights.

Cybersecurity is where Mistral is aiming the model most directly. The company says Large 4 scores 51 percent on the Artificial Analysis Cyber Index, and that in its own tests it refused 96.6 percent of malicious cyber requests. At the same time, Mistral is unusually candid about the other side of that coin: during the preview, it says, it saw a potential for uplift in conducting real attacks, and it is tightening its guardrails in response. A more aggressive variant built for penetration testing and red-teaming will not be public at all; it stays restricted to vetted partners, including Orange Cyberdefense, under contract. The message is that the same capabilities that make the model useful to defenders make it risky in the open, and Mistral is drawing the line at the most attack-oriented version rather than at the model itself.

The license draws a second line, this one commercial. Large 4 is released under the Mistral Ecosystem Open-Weight License, a modified MIT license that permits use, modification, distribution, and further training, including selling products built on the model. But a business that operates a model-as-a-service offering, renting access to the model or its derivatives to customers, must obtain a separate commercial license from Mistral first. That carve-out is the difference between this release and a fully open release: a company can deploy Large 4 internally or embed it in its own products, but it cannot simply resell hosted access to it without a deal. Developers can also still use the model through Mistral’s own API, where Mistral has cut prices to about half the preview rates: 0.70 dollars per million input tokens and 2.10 dollars per million output tokens, with cached input at 0.07 dollars. Microsoft Foundry is the first cloud to carry the model, and Nvidia offers it as a NIM container; self-hosted deployment is supported through the vLLM and SGLang serving frameworks.

The timing is part of the story. When Mistral unveiled Large 4 on October 6, it described it as the most capable open-weight model developed outside China, and promised downloadable weights by the end of October. It delivered on October 29, days before that deadline, and weeks after Chinese labs shipped their own large open-weight models through the summer and autumn. The result is a genuinely competitive open-weight frontier: Kimi K3.1 Pro, GLM-5.3, DeepSeek V4.1-Flash, and now Large 4, all downloadable, all with million-token context windows, and all priced well below the closed frontier models. For a European government or enterprise that wants a frontier model without depending on a US or Chinese provider, Large 4 is now the most concrete option on the table, and its sovereignty pitch, trained in Europe on 3,800 Nvidia Grace Blackwell GPUs in Mistral’s own data centers, is aimed squarely at that buyer.

That pitch has an unavoidable counterpoint, and Mistral does not hide it. Open weights cannot be recalled. Once the files are on Hugging Face, anyone can download them, and a determined user can try to fine-tune away the refusals that Mistral built in. The same property that protects a legitimate user from a vendor cutting off access also protects a malicious user from a vendor cutting off access. Whether Large 4’s defensive value outweighs its offensive risk is now a question that will be answered in the open, by the researchers, security teams, and attackers who download it, rather than by Mistral alone. The release is a bet that the balance favors openness, and that the teams defending networks will move faster with a trillion-parameter model in their own hands than the people trying to break into them.

What happens next is measurable. Independent benchmarks on the released weights will show whether Mistral’s coding and cyber scores hold up outside its own tests. The technical report, when it arrives, will show how the model was trained and evaluated. And the license’s hosted-service carve-out will be tested by the first company that tries to build a business on renting out Large 4. For now, the headline is simple: the model Mistral previewed in early October is no longer something you can only rent. It is something you can download, and that changes who gets to decide what it is used for.