Mistral Large 4 "Le Chonk" goes open-weight: a trillion parameters and 52B active

⚡ The gist in 5 seconds

• Mistral has open-sourced the weights of Mistral Large 4, known as Le Chonk: about a trillion parameters, roughly 52 billion active per token.
• The model runs on your own infrastructure: in the official vLLM example it is spread across eight GPUs.
• The license is the Mistral Ecosystem Open-Weight License, with separate terms for Model-as-a-Service.

🔍 What happened

Mistral has open-sourced the weights of Mistral Large 4 — a model the company calls Le Chonk. A preliminary version was shown on October 6, and the weights of the updated model were released on October 29, 2026 (the date is as given in the WIRED piece retold on Habr). Between these stages, reinforcement fine-tuning was underway: coding and cybersecurity tasks were made progressively harder.

The architecture is Mixture of Experts: for each token, a subset of specialized blocks is selected. About 52 billion parameters are active, counting input and output layers, so the trillion in the model card doesn't mean "a trillion in every computation." But the weights still have to fit entirely in memory — in the official vLLM example the model is distributed across eight GPUs. Downloading alone doesn't make running it cheap.

The leadership claims rest on three benchmarks. FrontierCode 2.0 from Cognition — 19.7%: a closed test that measures the quality of code fixes and the likelihood a project maintainer would accept them. Artificial Analysis Coding Agent Index — 48%, roughly on par with GLM-5.3 from Z.ai, but using on average 47% fewer tokens per task. Terminal-Bench 4.0 — 41%, up 13 percentage points from the preliminary version. The license is the Mistral Ecosystem Open-Weight License, a modified MIT with separate terms for the Model-as-a-Service business and an exemption for internal use.

🧠 Our take

Mistral's main argument isn't benchmarks but vendor lock-in. VP of Science Pierre Stock puts it this way: if a model is helping you work through an incident and the provider cuts your request limits mid-investigation, the team is left with a hole and no tool. Our assessment: this is the most honest argument in the release, because it's about procedure, not percentages.

What this changes in practice. For corporate security, open weights remove one risk and add another: the developer can't revoke access to what you've downloaded, but servers, electricity, the environment, and admins are now on your balance sheet. A 47% token saving per task looks nice, but the infrastructure bill is computed from hardware, utilization, and runtime — you can't carry a percentage from a benchmark into a project budget. In our experience, models like this pay off for those who already have a GPU fleet and a team to maintain it.

Who wins. Mistral gets a strong line in the conversation about a European alternative to Chinese models — and that's positioning, not measurement. The losers are those selling "just an API" without availability guarantees. And there's a second layer the release barely mentions: the same weights are available to defenders and attackers alike. Rafe Pilling of Sophos calls it dual-use technology, and the country of development makes no difference here.

⚠️ What we don't know

The weak spot is the numbers. The 19.7% on the closed FrontierCode can't be compared to anything without a public table, and the 48% on the Cognition index only makes sense alongside other runs of the same methodology. The dates in the piece (October 29, 2026) look odd and should be checked against the model card. We read a single retelling of a translation, not the primary model card and leaderboards, so treat the percentages as company-claimed, not independently verified. What exactly "fine-tuning between versions" means isn't disclosed in the source.