Mistral Large 4 is free to download, and it fits on one $22,299 Mac

Mistral has published the full weights of Mistral Large 4, the model it nicknames Le Chonk, so anyone can download it, modify it and run it on their own hardware. The French lab says it is now the strongest open model for writing code and for defensive security work, such as reproducing a software flaw and then patching it. The catch for anyone hoping to run it at home is size.

Le Chonk has about a trillion parameters, of which 49 billion are active for each word it produces. Even the most compressed versions, 2-bit builds made by the llama.cpp team and Unsloth, need a single Apple Silicon Ultra machine with 512 GB of unified memory, according to Hugging Face. The only Mac Apple sells with that much memory is a Mac Studio with the M5 Ultra chip, and 9to5Mac reports it costs $22,299 in the US configured that way. Every other way to run it means several high-end graphics cards working together.

Mistral’s case for the model rests on code. The company says it scores 41% on Terminal-Bench 4.0, a test of AI agents working in a command line, the best result among open models, and that it ties GLM-5.3 on Artificial Analysis’s coding-agent index while using 47% fewer tokens per task, which makes it cheaper to run. In blind side-by-side ratings from Surge, graders preferred its code to that of GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, and ranked only Anthropic‘s Claude Opus 5.5 above it.

The security pitch is aimed at defenders. Hugging Face says the model scored 82% on an Artificial Analysis task that asks a model to reproduce a real vulnerability in open-source software and then fix it, and that Mistral reports several closed frontier models score near zero there because they refuse to try. A bank or a government agency can run the model on its own servers, feed it its own code and crash logs, and keep using it if a provider changes its terms. “AI is going to be like electricity. You don’t want to live in a world in which your neighbor can cut off electricity from one day to another,” Pierre Stock, Mistral’s vice president of science, told reporters, according to Wired.

Most of those numbers are Mistral’s own, on benchmarks Mistral chose, and outside testing has barely started. On Artificial Analysis’s overall intelligence index, Hugging Face puts it fifth among open models, not first. The security argument also cuts both ways. Mistral wrote that while watching live traffic during the preview it saw “a potential for uplift to conduct real attacks” and is tightening its guardrails, but public weights can be downloaded and retuned by anyone, including attackers. “It’s clearly a dual-use technology,” Rafe Pilling, senior director of threat intelligence at Sophos, told Wired.

Home use is also a compromise even with the right Mac. Hugging Face ran the 2-bit build on an older M3 Ultra Mac Studio and got about 15 tokens a second, fine for chatting but slow for long coding jobs. On four Nvidia RTX 6000 Pro cards with very long documents, output dropped to about 3 tokens a second. Squeezing the model into 2 bits also costs accuracy compared with the full version Mistral benchmarked.

The weights are on Hugging Face for download worldwide under the Mistral Ecosystem Open-Weight License, a modified MIT license that allows free use, fine-tuning and commercial products, with one exception: companies that sell the model as a hosted service need a separate commercial license from Mistral. Everyone else can also use it through Mistral’s API at $0.70 per million input tokens and $2.10 per million output tokens, or through OpenRouter and Microsoft Foundry. Mistral released a preview earlier in October and published the final weights on October 29. Apple says Mac Studios ordered with 512 GB will ship between late November and early December.

“Once you open source the weights, they’re all over the internet. It’s impossible to cut off access,” Stock told Wired. For anyone without a $22,299 computer, the cut-off is the price of the memory.

Tags: Cybersecurity, Hugging Face, open-weight AI, Mac Studio, Le Chonk, Mistral AI