Mistral Just Gave Away Its Biggest Model. That Is the Point

On October 29, the French AI lab Mistral published the weights of Mistral Large 4, its trillion-parameter flagship model. Anyone can now download the model from Hugging Face and run it on their own hardware. The model is nicknamed “Le Chonk,” a joke about its size that turns out to be the whole story: this is one of the largest models ever handed out for anyone to run.

The release matters less for any single benchmark score than for what it changes about control. With a closed model, you rent access. The provider can change the price, change the rules, or cut you off. With open weights, the model is a file you hold. That difference is why Mistral framed the release around sovereignty, and why governments and large companies, not just hobbyists, are the real audience.

What Mistral actually released

Mistral Large 4 first appeared as a public preview on October 6, available only through Mistral’s API. The October 29 release is an improved checkpoint: after three more weeks of reinforcement learning and feedback from preview users and security partners, Mistral published the full weights.

The headline numbers, from Mistral’s own documentation, are about 1.05 trillion total parameters, with roughly 49 billion active for each token, or 52 billion if you count the embedding and output layers as well. It is a mixture-of-experts design, which means only a slice of the model works on any given token rather than the whole trillion parameters at once. That is how a model this large stays affordable to run at all.

It is natively multimodal, taking text and images as input and producing text and code. Its context window is one million tokens, enough to hold a very large codebase or a long set of documents in a single session. Mistral says it was trained on 3,800 Nvidia Grace Blackwell GPUs in its European data centers, and that it supports more than 160 languages.

Alongside the weights, Mistral cut its API price to $0.70 per million input tokens and $2.10 per million output tokens, and made the model available through Mistral Studio, OpenRouter, Nvidia-hosted endpoints, and Microsoft Foundry, which is the first cloud to carry it for private deployment.

Open weights are not the same as open source

Mistral calls this an open-weight release, and the distinction matters. The weights are published under the Mistral Ecosystem Open-Weight License, a modified MIT license. You can use the model, modify it, fine-tune it, and build commercial products on it for free.

There is one important carve-out: a business that offers the model itself as a hosted service, a model-as-a-service business, needs a separate commercial license from Mistral. In plain terms, you can build a product powered by Large 4, but you cannot simply resell Large 4 as your own hosted model service without talking to Mistral first. That clause is where “open” stops and Mistral’s business model begins, and any company planning to build on the model should read it before committing.

The other half of the distinction is practical. Open weights cannot be recalled. Once the files are downloaded, Mistral cannot push an update that changes what every copy does, and it cannot revoke access the way an API provider can. Pierre Stock, Mistral’s vice president of science, put it directly to TechCrunch: open weights give organizations access “that can’t be cut off by definition.” That permanence is the feature for defenders and the risk for everyone else, and both sides of that sentence are true at the same time.

The numbers, and how to read them

Mistral’s reported results are strong on the two jobs it is pitching hardest: coding and cybersecurity. On Terminal-Bench 4.0, a test of real terminal-based software work, Large 4 scores 41 percent, up 13 points from the preview. On Artificial Analysis’s Coding Agent Index it scores 48 percent, effectively tied with China’s GLM-5.3, the leading open model on that test, while using 47 percent fewer tokens per task. On cybersecurity, Mistral reports 93 percent on Cybench, a set of security challenges, and 82 percent on reproducing and patching a real vulnerability in open-source software. It also says the model refused 96.6 percent of malicious cyber requests in its tests.

Independent scoring tells a more mixed story, which is normal for a release this fresh. Artificial Analysis’s October 29 assessment gives Large 4 a score of 26 on its Intelligence Index at the high reasoning setting, up from 25 for the preview. That puts it above Meta’s Muse Spark Open, making it the highest-scoring model with available weights developed outside China in that evaluation, but still well below DeepSeek and GLM-5.3 at 32, and far below the leading closed models. The honest reading is that Large 4 is the best open-weight coding and cyber model outside China by some measures, and not the best model overall by others.

The more interesting number is economic. Artificial Analysis reports the cost per Intelligence Index task fell to $2.06, down from $5.54 for the preview, a 63 percent reduction. That is the number that decides whether a model gets used in production, and it is the number Mistral’s API price cut is designed to reinforce.

All of these figures should be read with one caveat attached: they are Mistral’s numbers, or measurements of Mistral’s checkpoint by evaluation firms, not yet independent reproductions by outside labs running the model themselves. Mistral says a full technical report is coming soon. Until then, treat the scores as a claim to test, not a verdict.

The other model launch this week took the opposite path

Large 4 did not arrive in a vacuum. One day earlier, on October 28, Google unveiled Gemini 4 Carbon, a frontier model also pitched at cybersecurity, and limited first access to a subset of vetted defenders through its Fairwind Program while it tests safeguards. Same week, same pitch, opposite access philosophies: Google keeps its most capable cyber model behind a gate; Mistral hands its model to everyone.

That contrast is the real story of the week. Google’s argument is that a powerful cyber model in the wrong hands makes attacks easier, so access should be earned. Mistral’s argument is that defenders are already outgunned, and that keeping the best defensive tools locked up leaves everyone except the largest companies exposed. Both arguments have merit, and the industry is now running both experiments in public at the same time.

Mistral is unusually candid about the risk in its own case. It says it saw “a potential for uplift to conduct real attacks” during the preview, and it is tightening guardrails in response. A more aggressive penetration-testing version of the model stays restricted to vetted partners under contract. That honesty is welcome, but it does not solve the core problem: the downloadable weights can be modified, and no license stops a determined user from trying.

Why sovereignty is selling right now

The word Mistral keeps using is sovereignty, and it is landing because the buyers are worried about something specific. A European bank, a government agency, or a manufacturer with trade secrets does not just want a good model. It wants to know where its data goes, which country’s rules govern it, and whether access can disappear because of a policy change in another capital.

Self-hosting answers those questions in a way no API contract fully can. Data stays inside the organization’s own systems. The model can be fine-tuned on private code and documents without sending them to a third party. And no provider can change the model underneath a deployed product overnight. For regulated industries, that control is often worth more than a few extra benchmark points.

Europe’s angle is geopolitical as well. Mistral is the continent’s flagship lab, and Large 4 is its proof that a European company can train a trillion-parameter model on European infrastructure and give it away on terms that favor European users. The open-weight race is now also a contest over who gets to set the defaults: American closed models, Chinese open models, or a European alternative. Large 4 is Mistral’s bid for that third lane.

The hard parts nobody should skip

None of this makes Large 4 easy to run. The full-precision checkpoint is about 1,061 gigabytes, and the compressed NVFP4 version is about 610 gigabytes. Per the SGLang inference documentation, the standard version needs eight Nvidia B300 GPUs, with the compressed variant running on four GB300 GPUs. That is data-center hardware, not a laptop project. Most organizations that “self-host” Large 4 will in practice run it through a cloud partner like Microsoft Foundry or Nvidia, which brings back some of the dependence self-hosting was meant to remove, even if the model file itself remains theirs.

There is also an operations gap. Running a frontier model means monitoring it, updating it, patching its serving stack, and watching what it does with tools and data. The October agent incidents, from Anthropic’s fake police tip disclosure to the Pentagon contractor story this week, all point the same way: the danger is rarely the model alone. It is the model plus credentials plus tools plus nobody watching the logs. An open-weight model in your own data center does not remove that job. It moves it onto your team.

And there is a question the license does not answer cleanly: where does building a product on Large 4 end and operating a model-as-a-service begin? Companies that wrap the model in an API for their own customers should get that answer in writing before they scale, because the free license has a boundary and Mistral drew it deliberately.

What builders should do this week

First, test before you commit. The API route, at $0.70 per million input tokens, is the cheap way to find out whether Large 4 actually beats what you use now on your own code and your own security tasks. Benchmarks are a starting point, not a substitute for your own test set.

Second, if you self-host, start with a narrow job. Mistral says customers can tune Large 4 into smaller, task-specific variants for well-defined security work. A focused model for vulnerability triage or patch review is a far better first deployment than a general assistant for everything.

Third, treat the cyber pitch as two-sided from day one. The same capability that finds vulnerabilities in your code can find them in someone else’s. Log tool calls, restrict what the model can touch, and keep a human in the loop for anything that changes a live system.

Fourth, read the license and the deployment docs, not just the announcement. The model-as-a-service clause, the checkpoint sizes, and the hardware requirements are the parts that decide whether this model fits your budget and your legal position.

What to watch next

Three things will decide whether Large 4 is a milestone or a moment. The first is independent reproduction: whether outside labs running the downloaded weights get results close to Mistral’s claims. The second is adoption: whether enterprises and governments actually deploy it, or whether most usage stays on Mistral’s API and partner clouds. The third is the next open release: Moonshot’s Kimi K3.1 Pro, whose weights are due by November 4 and which already leads Terminal-Bench 4.0 at 67 percent in Mistral’s own comparisons, will reset the open-weight leaderboard if it ships on time.

Mistral gave away its biggest model because it believes openness is how a smaller lab beats larger rivals. The bet is that control, not just capability, is what buyers want most right now. This week gave that bet its first real test, and the results will show up not in benchmark tables but in who actually runs the model, where, and for what.

Sources

• TechCrunch, “Mistral releases weights for its trillion-parameter AI model, targeting cybersecurity,” October 29, 2026.
• VentureBeat, “Mistral Large 4 ‘Le Chonk’ arrives as the top open model outside China with improved coding performance and 63% lower cost-per-task,” October 29, 2026.
• Artificial Analysis, October 29, 2026 assessment of Mistral Large 4 (Intelligence Index v5, Coding Agent Index, Cyber Index).
• Mistral AI announcement and Hugging Face model listing for Mistral Large 4, October 29, 2026.
• Google Gemini 4 Carbon coverage, October 28, 2026, for the Fairwind Program comparison.

Related reading

Google Built Gemini 4 Carbon to Fight Hackers | Google Limits Free Gemini Starting Today. Here Is What Actually Changes | Claude Filed a Fake Police Tip. Anthropic Reported It First