Mistral released the weights of Mistral Large 4 on Thursday, turning its flagship preview into something anyone with the hardware can download, inspect and run. The model, nicknamed "Le Chonk," is a natively multimodal system with roughly one trillion total parameters, of which about 52 billion activate per token. It is on Hugging Face in full and quantised builds, the latter sized for four Nvidia B200 or B300 GPUs.
The interesting part is not the parameter count — open-weight releases are now a monthly event — but the delta since the preview three weeks ago. Mistral says continued reinforcement learning, guided by early customers, lifted Terminal-Bench 4.0 from 28% to 41%, putting it just ahead of GLM-5.3 at 40% and ahead of every other open-weight model on that test. On the Artificial Analysis Coding Agent Index it scores 48%, effectively tied with the leading open model while burning 47% fewer tokens per task. Output tokens on terminal tasks fell by 30%. On Cognition's private FrontierCode 2.0 benchmark, which asks whether a maintainer would actually merge the agent's patch, it leads at 19.7%.
That last number is the one to sit with. A 19.7% merge-worthy rate is a frontier claim that still implies four out of five agent patches are not good enough to ship — the sort of honesty benchmark tables usually sand off.
Mistral also leaned hard into defensive security. Mistral Large 4 scores 79% on CyberGym-E2E, which tests whether a model can find, reproduce and patch real memory-safety bugs in open-source C and C++ projects — level with MiMo-V2.6-Pro and ahead of its Chinese open-weight rivals. It manages that while refusing 96.6% of malicious cyber prompts across JailbreakBench, StrongREJECT and AgentHarm, which Mistral calls the highest refusal rate it measures among open models. The company concedes the trade-off is live, not theoretical: after watching traffic since the preview, it says it saw "potential for uplift to conduct real attacks" and is adjusting guardrails. A tougher variant for penetration testing goes only to vetted design partners such as Orange Cyberdefense, under contracts restricting testing to systems the customer owns.
The scarcity argument is where Mistral wants the conversation. The model was trained from scratch on European infrastructure using 3,800 Nvidia Grace Blackwell GPUs; the reinforcement learning ran on Mistral's own compute in France, drawing about 10 megawatts. "Other labs have raised hundreds of billions to develop equivalent capabilities; in three years, Mistral raised roughly $6 billion," CEO Arthur Mensch said. Pierre Stock, the company's VP of Science, framed open weights as an insurance policy: AI will be like electricity, and no company should depend on someone who can cut off access "from one day to another."
The licence is where that rhetoric meets a lawyer. Mistral Large 4 ships under the Mistral Ecosystem Open-Weight licence, a modified MIT that allows use, modification, fine-tuning and commercialisation — with one carve-out: anyone running it as a model-as-a-service offering needs a separate commercial licence from Mistral. That is open weights, not open source, and it keeps the hosted-inference business on Mistral's side of the fence. API pricing is $0.70 per million input tokens, $0.07 cached, $2.10 output. A quantised build runs on four B200/B300 GPUs; smaller distilled versions are promised, including one for a laptop, without dates.
Scepticism is warranted on two fronts. Every benchmark above is either vendor-run or drawn from an index the vendor chose to publish, and FrontierCode 2.0 is a private test set — a strong signal, not an audited one. "Best open-weight model for coding" is a claim that will be re-tested by independent evaluators within days, and the gap to closed frontier models on human preference remains: in Surge Code's blind pairwise grading it ranked behind only one closed model, but behind it nonetheless.
The larger point stands regardless. A trillion-parameter model with frontier-adjacent coding and security scores is now something an enterprise can put on its own metal, in its own jurisdiction, with no API dependency and no one able to revoke it. Three weeks of reinforcement learning moved it from 28% to 41% on the benchmark that matters most to agent builders. That cadence, not the weights themselves, is the thing to watch.
Why it matters: the open-weight tier has been Chinese-led for a year. Mistral is now arguing that a European lab with roughly a twentieth of the capital can hold the same line — and that the real product being sold is not intelligence but the ability to keep it running after someone else's policy changes.
