Mistral Large 4 open weights landed on October 29, 2026, and they change who gets to run a frontier-class model. France’s Mistral AI published the full weights for Large 4, nicknamed Le Chonk, on Hugging Face three weeks after its October 6 preview. Anyone can now download it and run it on their own servers.

Don’t expect to run it on a laptop. Large 4 holds about 1 trillion parameters in total, with roughly 49 billion active for each token, and it takes text and images with a 1-million-token context window. Through Mistral’s own API, it now costs $0.70 per million input tokens and $2.10 per million output tokens, about half the preview price of $1.36 and $4.18.

That’s the bet.

Mistral is pitching Large 4 first to cyber defenders, not to chatbot users. In its announcement, the company said enterprises and governments can now “download, deploy, and own frontier AI to detect, analyze, and respond to threats across code and IT infrastructure.” Security teams, banks, and hospitals can keep code, logs, and incident data inside their own firewall instead of sending it to someone else’s API.

What Mistral Large 4 open weights actually give you

Open weights are not the same as open source. You get the trained model files, so you can run the model, adapt it, and keep it working even if a vendor changes its terms or pulls an API. You don’t get Mistral’s training data or its full training recipe.

What you do get is a mixture-of-experts design that only wakes up a small slice of itself for each token, which keeps the cost per token down even though the total model is huge. Self-hosting a model this size still takes serious GPU capacity, plus the engineering work to serve it safely.

The license catch

Read the license before you build a business on it. Mistral calls the terms its Ecosystem Open-Weight license, a modified MIT license, which lets you use, modify, fine-tune, and sell products built on the model for free. One carve-out matters for cloud providers: a business offering the model itself as a hosted service must get a separate commercial license from Mistral first.

A bank running Large 4 inside its own systems is in a different position from a startup renting Large 4 access to the public by the token, and Mistral has not spelled out where a normal product wrapping the model ends and a model-as-a-service business begins. Get that answer in writing before you launch.

The cyber defense pitch, and the risk Mistral admits

Mistral’s numbers, all from the company rather than independent labs, put Large 4 at 48% on the Artificial Analysis Coding Agent Index, effectively tied with China’s GLM-5.3 while using 47% fewer tokens per task. Its Terminal-Bench 4.0 score rose from 28% at preview to 41%, Mistral says, and it reports 51% on the Artificial Analysis Cyber Index with a 96.6% refusal rate on malicious cyber prompts in its own tests.

Pierre Stock, Mistral’s vice president of science, told TechCrunch that open weights give defenders something a hosted API can’t: “Access that can’t be cut off by definition because the model is open-weighted.” If a provider changes policy or goes down, your defense tool keeps running.

Mistral also said something labs rarely put in a launch post: during the preview, it saw “a potential for uplift to conduct real attacks” in live traffic, and it says it is tightening guardrails in response. A stronger variant built for penetration testing stays limited to vetted partners on authorized systems. Wired’s coverage made the same point from the other side, with a Sophos executive warning that the freedom to download and customize a model helps attackers as well as defenders.

Open weights can’t be recalled. Once the files are out, a determined user can try to strip out the refusals, and no license update will pull the download back. Mistral’s answer is that defenders need the capability fast because attackers will get good models anyway, a view CEO Arthur Mensch summed up by saying AI should “be a tool for empowerment, not dominance.”

Why this matters

Google took the opposite route two days earlier: its Gemini 4 Carbon model, announced October 28, goes first to selected cyber defenders through Google’s Fairwind program, with no public date for wider access. Mistral handed everyone the files instead.

For US readers, the practical question is whether your team can actually use an open model safely. A security team with its own infrastructure can test Large 4 on its own code and keep the results private. A smaller team will likely use the API or a cloud host instead, which brings back some of the dependence open weights were supposed to remove. Watch next for independent reruns of Mistral’s benchmarks, and for how Mistral polices its model-as-a-service carve-out.

Frequently asked questions

What are Mistral Large 4 open weights?

Open weights are trained model files anyone can download and run privately. Mistral published Large 4 on Hugging Face on October 29, 2026, so teams can host it instead of calling Mistral’s API.

Can anyone use Mistral Large 4 for free?

The weights are free to download, use, and modify under Mistral’s modified MIT license. Companies offering the model itself as a hosted service need a separate commercial license.

How big is Mistral Large 4, and what does it cost through the API?

It has about 1 trillion parameters in total, with roughly 49 billion active per token. Through Mistral’s API it costs $0.70 per million input tokens and $2.10 per million output tokens, about half the preview price.

Is Mistral Large 4 actually the best open model for coding?

Mistral says it ties China’s GLM-5.3 at 48% on the Artificial Analysis Coding Agent Index while using 47% fewer tokens per task. Those are Mistral’s numbers, not independent results.

Sources: TechCrunch, Wired, Mistral AI