This week: Mistral released the open weights of Mistral Large 4, a 1-trillion-parameter model it nicknames Le Chonk. This means anyone with enough hardware can download, run, and fine-tune it.

Elsewhere: What actually changed since the preview, how to try it without a server room, and the one licensing condition worth reading twice.

The New Prime Signal

Most AI companies name their flagship models like luxury sedans. Mistral named its biggest one Le Chonk. On October 29, the French lab released the weights of Mistral Large 4, the 1-trillion-parameter model it first showed as a preview earlier this month. That means you can download it, run it on your own machines, and fine-tune it, instead of renting it through someone else's API. You will need a lot of machine. It is called Le Chonk for a reason.

What happened: Mistral says the open release is an improved version of the preview, after about three more weeks of reinforcement learning. According to the model card, it has roughly 1 trillion parameters in total, but only about 49 billion are active for each token (52 billion counting embeddings and output layers). It reads text and images, writes text, handles a context window of 1 million tokens, and Mistral says it is fluent in more than 160 languages. Mistral trained it from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own data centers in Europe. It also comes with three reasoning modes—low, medium, and high—so you can trade thinking time for speed and cost.

Why it matters: Open models this large are rare, and ones built in Europe are rarer. For a bank, a hospital, or a government office that can't send its data to someone else's servers, being able to host a frontier-class model in-house is the whole point. Mistral says it is now the best open-weight model for coding. On the Artificial Analysis Coding Agent Index, the company reports 48%, which it calls "effectively tied" with the top open model, GLM-5.3, while using 47% fewer tokens per task. It also says the model improved by 13 points on Terminal-Bench 4.0 since the preview, reaching 41%. Fewer tokens for the same work is the kind of improvement your cloud bill notices.

The fine print: Every benchmark above comes from Mistral itself, and some of the same scores appear with different numbers in Hugging Face's launch post, so check which version a chart describes. On its own model card, Mistral admits the leading open-weight models from China still score higher on the broad Artificial Analysis Intelligence Index. The license is a modified MIT license: free to use, modify, fine-tune, and sell products with, but a business that runs a model-as-a-service offering needs a separate commercial license from Mistral. Mistral also says that, while watching live traffic since the preview, it saw "a potential for uplift to conduct real attacks," and is tightening its guardrails. And "open" does not mean "runs on your laptop."

This is a real gift to anyone who wants serious AI without sending their data off to a black box. A trillion-parameter model with downloadable weights, clear training origins, and a cheaper low-effort mode is a meaningful step for open AI outside the U.S. and China. The scoreboard claims deserve independent testing, which Mistral is openly inviting. Also, any lab that ships a trillion-parameter model and calls it Le Chonk gets bonus points from us.

The trick is that Le Chonk is huge on paper but only wakes up a small slice of itself for each word.

How it works: Mistral Large 4 is a mixture-of-experts model. Inside are 128 specialist "experts," and for every token a router picks 4 of them, plus one shared expert that always runs. Hugging Face's technical walkthrough describes a second shortcut: a sparse-attention indexer that picks which earlier parts of the text to look at, instead of rereading everything. Together they keep a trillion-parameter model from needing a trillion parameters' worth of work per word. For images, it uses a 1.6-billion-parameter Pixtral vision encoder. The checkpoint also ships with a small built-in draft model called DSpark that guesses a few tokens ahead, and Mistral says it measured up to 1.9 times higher throughput with it turned on, without changing the answers.

Try it: The easiest route needs no hardware at all: Mistral Studio, the Mistral API, and OpenRouter all serve it, with API pricing of $0.70 per million input tokens and $2.10 per million output tokens. The model card also links the Le Chonk Gallery of demo apps, which Mistral says run on a small shared free budget per visitor during the release's first week. If you want it at home, Hugging Face reports the most compressed 2-bit version runs on a single Apple Silicon Ultra machine with 512 GB of memory, at about 15 tokens per second. Everyone else will want a rack of GPUs, vLLM or SGLang, and a good reason. Start with a coding task you already know the answer to, and check its work.