A trillion parameters. Free to download.

Yesterday, Mistral released the full weights of Mistral Large 4, a one-trillion-parameter model the company nicknamed "Le Chonk." Anyone can pull it from Hugging Face, run it on their own servers, fine-tune it, and build a business on it.

That makes it one of the largest open-weight models ever released, and Mistral says it is the strongest open model on agentic coding. Through the API it costs $0.70 per million input tokens and $2.10 per million output, roughly a tenth of what Claude Opus 5.5 charges for output.

It is not the best model in the world. Claude and Gemini still lead by a wide margin. But it is the best model you can actually own, and Mistral built its whole pitch around that: if the weights sit on your hardware, nobody can take them away.

Here is what Mistral shipped, where it stands against the closed labs, and why its bet on cybersecurity matters.

One trillion parameters, about 52 billion active. Mistral Large 4 is a Mixture-of-Experts model, so only a slice of it runs for each token. The Hugging Face repo name points to roughly 52 billion active parameters, which is what keeps it affordable to serve.

One model for everything. It combines instruction following, reasoning, and agent work in a single natively multimodal model, speaks more than 160 languages, and has three reasoning modes: low for simple tasks, medium for everyday work, and high, the default, for the hardest problems.

Actually open. The weights ship under Mistral's Ecosystem Open-Weight license, a modified MIT license. You can use, modify, fine-tune, and sell products built on it. The one catch: companies that resell it as a hosted model service need a separate commercial license.

Runs almost anywhere. It is on Mistral's API, OpenRouter, Microsoft Foundry, and NVIDIA's hosted endpoints, and you can deploy the weights yourself with vLLM or SGLang. A full technical report is still "coming soon."

Mistral's numbers make a clear case for best open model, and an honest one for not-quite-frontier.

Coding. Mistral Large 4 scores 48 on the Artificial Analysis Coding Agent Index, tied with China's GLM-5.3 at the top of the open models, while using 47% fewer tokens per task. Claude Opus 5.5 scores 62 and Gemini 4 Carbon 59. On Cognition's private FrontierCode 2.0 it scores 19.7%, the best open-weight result shown.

Agents. It scores 41% on Terminal-Bench 4.0, up from 28% in the preview, and 84.5% on BrowseComp for web research. In Surge's human preference tests for code, raters preferred it over GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, ranking it behind only Claude Opus 5.5.

Work tasks. It hits 73.3% on FinWorkBench, second among open models only to Kimi K3.

The honest read. These are mostly Mistral's own numbers, and the gap to the closed frontier is real: about 14 points behind Opus 5.5 on coding agents. What you get in exchange is control and price.

Mistral is pitching Large 4 hardest at security teams, and that is the most interesting part of the launch.

The model is built to detect, analyze, and respond to threats across code and IT infrastructure, entirely on-premises. For banks, governments, and defense contractors that cannot send code to an American cloud, that is the whole point. "It can't be cut off," Mistral VP of Science Pierre Stock told TechCrunch.

Strong guardrails, by open-model standards. Mistral reports a 96.6% average refusal rate on malicious cyber requests, the highest it measured among open models, and 97.1% attack resistance on the B3 Agent Security Benchmark. It scores 51% on the Artificial Analysis Cyber Index.

And an admission. Mistral says it saw the model refuse attacks on live targets, but also saw "potential for uplift" to real attacks, so it is tightening its guardrails. A stronger cybersecurity-specialized version is available only to vetted enterprise partners, for testing on systems they own.

Compare that with Google, which this week held Gemini 4 Carbon back for cyber defenders only. Mistral just put a capable cyber model on Hugging Face for anyone. Open weights cannot be recalled, and that is the tradeoff the whole industry is now arguing about.

Open models are no longer a hobbyist side show. A trillion-parameter model that anyone can download sits within striking distance of the closed frontier, at a fraction of the price.

The economics are hard to ignore. At $2.10 per million output tokens, Mistral Large 4 is about a tenth of Opus 5.5's output price and about a fifth of Gemini 4 Carbon's introductory rate. For high-volume agent work that does not need the very best model, that gap pays for a lot of engineering.

Four things to do with this:

Benchmark it on your own workload. For bulk coding, document, and support tasks, it may be good enough at a tenth of the cost.

Keep the frontier for the hard stuff. For your most difficult agentic coding, Opus 5.5 and Gemini 4 Carbon still lead by double digits.

If data can't leave your building, look closely. On-premises deployment is the real selling point for regulated industries.

Read the license. Free for your own products, but hosting it as a paid model service needs a commercial deal with Mistral.

The stories moving fast beyond today's headlines:

A trillion parameters, open. Mistral released the full weights of Mistral Large 4, a 1-trillion-parameter Mixture-of-Experts model with about 52 billion active parameters, under a modified MIT license.

Best open model, not best model. It ties GLM-5.3 at 48 on the Artificial Analysis Coding Agent Index with 47% fewer tokens, but trails Claude Opus 5.5 (62) and Gemini 4 Carbon (59).

Dirt cheap. API pricing is $0.70 input and $2.10 output per million tokens, about a tenth of Opus 5.5's output price.

Built for security teams. It targets on-premises threat detection and response, with a 96.6% refusal rate on malicious cyber requests, while Mistral admits it is still tightening guardrails.

What to actually do. Test it on your high-volume workloads, keep closed frontier models for the hardest tasks, and check the license before you host it for others.

Stay building. 🤖