Mistral releases open weights for its Large 4 flagship model

Mistral AI on Thursday released the weights of Mistral Large 4, the 1-trillion-parameter, natively multimodal model it put into public preview earlier this month. The company says three weeks of additional reinforcement learning since the preview made it “the best open-weight model for coding.” Every score below is one Mistral reported; a full technical report is “coming soon,” the company says.

What it does: On Terminal-Bench 4.0, Mistral says the model improved by 13 percentage points from its preview to 41%, the lead among open-weight models, while cutting output tokens by 30%. On the Artificial Analysis Coding Agent Index it scores 48%, which Mistral calls “effectively tied” with GLM-5.3 while using 47% fewer tokens per task on average. It leads Cognition’s private FrontierCode 2.0 benchmark at 19.7%, the company says, and in Surge Code’s blind human-preference tests it ranked behind only Claude Opus 5.5 among the models evaluated. The model now offers three reasoning modes: Low, Medium and High, the default.

The cyber question: Mistral says the model scores 51% on the Artificial Analysis Cyber Index, which tests finding, reproducing and patching vulnerabilities, and refused 96.6% of malicious cyber requests on average across the cyber-related prompts of JailbreakBench, StrongREJECT and AgentHarm. The post also says that while monitoring live traffic since the preview, “we see a potential for uplift to conduct real attacks,” and that Mistral is modifying its guardrails and will keep updating them. Separately, enterprise design partners such as Orange Cyberdefense get preview access to a specialized cyber variant for penetration testing and red-teaming, subject to “rigorous vetting and contractual safeguards.”

Other results: Mistral reports 73.3% on the Finch finance benchmark, 84.5% on BrowseComp and 97.1% attack resistance on the B3 Agent Security Benchmark, up from 93.3% in preview.

Who can use it: The weights are on Hugging Face under the Mistral Ecosystem Open-Weight license, a modified MIT license. It is free to use, modify and commercialize, but businesses running a Model-as-a-Service offering need a separate commercial license. Through the Mistral API it costs $0.70 per million input tokens, $0.07 cached and $2.10 for output. It is also on OpenRouter, Nvidia-hosted endpoints and Microsoft Foundry, which Tina Schuchman, a Microsoft corporate vice president, said makes Microsoft the first cloud provider to offer it. Executives from Booz Allen Hamilton, Cisco, Cloudflare and Hugging Face are quoted backing the release.

Why it matters: The rare part of this release is the warning built into it. Mistral is handing anyone the weights of a model that it says could give attackers uplift, and once weights are downloaded, the guardrail changes it promises can’t follow every copy. Its answer is to split access: the general model goes out openly, while the variant tuned for offensive testing stays behind vetting and contracts. That is close to the reverse of Google’s approach with Gemini 4 Carbon, which starts with cyber defenders before a wider rollout. For buyers, the pitch is control. The banks and security firms quoted stress running the model on their own infrastructure, and the license lets most of them do that without paying Mistral. The open question is the benchmarks, all reported by Mistral, until its technical report and outside testers weigh in.

What’s next: Mistral says Large 4 will be the base for a new generation of specialized models, and its technical report is still to come.