Mistral Large 4: A Trillion-Parameter MoE, Open Weights by Month's End
October 7, 2026 · 5 min read
On October 6, France's Mistral AI showed up in Abu Dhabi and unveiled Large 4 — a trillion-parameter mixture-of-experts model, codenamed "Le Chonk," roughly "the chubby." Only 49 billion parameters fire on any given token, so it runs less like a lumbering giant and more like a nimble fleet of specialists. It is natively multimodal, and it was trained in Mistral's own European data centers on 3,800 NVIDIA Grace Blackwell chips. (via Le Monde)
Read that again. A trillion parameters, built in Europe, launched in Abu Dhabi, trained on hardware sitting in Mistral's own datacenters — not rented from a US cloud giant. The company says Large 4 is the best open model on aggregated benchmark leaderboards in Europe and the US, and CEO Artur Mensch made the comparison explicit: on cybersecurity evaluations, he claims it beats the top Chinese models. Europe has spent years hearing it lost the AI race. This is the counterargument, arriving in a trench coat.
The rollout is the real story
The API preview goes live the same day. The weights don't — not until the end of October, and even then there is a detour: Mistral is handing red-team versions to cybersecurity experts and governments first, for early access before the public release. That is a sequencing choice you don't make by accident. Ship the model to the people paid to break it, let them kick the tires for a few weeks, then open the weights to everyone. It's a bet that the only way to release a trillion-parameter open model responsibly is to let the professionals try to weaponize it first. Or, less charitably: a few extra weeks to watch what breaks before the rest of us get the keys. Either way, this is Mistral saying the quiet part out loud — at this scale, "open weights" is no longer a move you make lightly.
One trillion in name, 49 billion at work
The headline number is one trillion parameters, but the working number is 49 billion — that's how much of the model activates per token. The other experts sit idle until they are called. That is the whole MoE pitch: frontier-scale knowledge, with the inference-time cost of a model a fraction of the size. And the training footprint matters as much as the architecture: 3,800 Grace Blackwell chips in European data centers. Sovereignty is part of the product here — this isn't just a model release, it's a statement that Europe can train at the frontier on its own soil. Nobody outside Mistral has independently verified the benchmarks yet, so the "best open model in Europe and the US" claim is exactly what it is: a launch-day claim. Treat it accordingly.
The China framing
Mensch didn't dance around it. On cybersecurity, he said outright that Large 4 outperforms the leading Chinese models — an unusual thing for a CEO to volunteer on stage, and a telling one. The benchmark isn't the lab anymore; it's Beijing, and the pitch is that open weights plus European infrastructure is a safer bet than closed Chinese systems. Notice where this was announced: Abu Dhabi, not Paris. The Gulf is where the compute money and the sovereign-AI contracts live right now, and Mistral flew the flagship there on purpose. The message to that audience is simple: if you want frontier AI that isn't American closed-source or Chinese closed-source, there is now a European open-weights option with a trillion-parameter receipt.
What to watch
Three things. First, the end-of-October weight release — promises of open weights have a way of slipping, and the red-team preview gives Mistral a ready-made excuse if this one does. Second, independent benchmarks: aggregated leaderboard claims are only as good as the aggregation, and the cybersecurity-versus-China claim needs third-party eyes. Third, that early-access guest list: if governments really do get hands-on first, watch which ones say yes. The list will tell you more about the model's reception than any leaderboard. One caveat to keep in mind: this is day-one reporting. The model that ships in the API preview and the weights that land at month's end are the facts that count. "Le Chonk" is a great name. Names are cheap; trillion-parameter weight files are not.