Le Chonk: Built for Coders, Trusted by Defenders.

Mistral AI has published the weights of Mistral Large 4, codenamed Le Chonk, a 1-trillion-parameter natively multimodal model that had been in public preview earlier in the month. The company says three additional weeks of reinforcement learning since the preview drove gains in coding and defensive cybersecurity, and that it is now the best open-weight model for coding. Mistral describes the model as trained from scratch on European infrastructure, with the release pitched at organizations that want to host, inspect and customize the model themselves rather than depend on an API served by Mistral.
The headline coding numbers are specific. On Cognition's private FrontierCode 2.0 benchmark, which judges end-to-end code quality and whether a repository maintainer would merge an agent's change, Mistral Large 4 scores 19.7%. On the Artificial Analysis Coding Agent Index it reaches 48%, described as effectively tied with GLM-5.3, the leading open-weight model there, while using 47% fewer tokens per task on average. On Terminal-Bench 4.0 it climbed 13 percentage points over the preview to 41%, taking the open-weight lead while cutting output tokens by 30%.
Mistral also cites human preference evidence. In blind pairwise evaluations run by Surge Code, the model was preferred over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, and ranked behind only Claude Opus 5.5 among the models evaluated. That kind of grader study complements automated benchmarks because it captures readability and judgment that unit-style tests miss, though pairwise preference does not by itself prove a change is correct or mergeable.
A practical addition is a set of three reasoning modes, which let callers trade latency and cost against depth of deliberation. Low is intended for simple tasks and replaces the older no-reasoning option, Medium for everyday work, and High, the default, for the hardest cases needing precision. For product teams this is effectively a pricing lever: routing routine requests to the low setting and reserving high-effort reasoning for hard problems should cut spend without changing infrastructure. This is my reading, but it is the obvious reason to expose explicit effort tiers.
On security, Mistral positions the model for defenders. It scores 51% on the Artificial Analysis Cyber Index, which tests finding vulnerabilities in a codebase, reproducing and validating them, and patching them without breaking existing functionality. Mistral says the model handles these legitimate defensive tasks without unnecessary refusals, and that its guardrails distinguish defensive from malicious cybersecurity work while refusing an average of 96.6% of malicious cyber requests across JailbreakBench, StrongREJECT and AgentHarm prompts, which it calls the highest refusal rate it measures among open models. Mistral also says it observed the model refusing black-box attacks on live targets but sees potential for uplift toward real attacks, and is adjusting guardrails in response.
Enterprise design partners, including Orange Cyberdefense, get preview access to a variant tuned for penetration testing, vulnerability assessment and red-teaming, gated by vetting and contracts restricting testing to owned or authorized systems. Cybersecurity vendors quoted in the announcement include Booz Allen Hamilton, Cisco, Cloudflare, CrowdStrike, Elastic, Fortinet, Rubrik, Zscaler and ENISA, and several emphasize self-hosting and data residency rather than raw scores.
Broader enterprise metrics also moved. Finch, a finance workflow benchmark covering spreadsheet creation and editing, reached 73.3%, behind only Kimi K3 among open models. BrowseComp hit 84.5%, up from a supervised fine-tuning checkpoint at 75%, above GLM-5.3 and DeepSeek Flash 4.1 1017. On the B3 Agent Security Benchmark, resistance to attacks such as system prompt exfiltration and malicious code injection rose to 97.1% from 93.3% in preview, which Mistral frames as meaning the model can host safe agents without extra tuning. Customers can customize it through Mistral Forge and build and govern agents with Mistral Studio, with financial references from La Banque Postale, ABN AMRO and Belfius.
A trillion parameters is far larger than typical open-weight releases, which usually sit in the tens to low hundreds of billions. In practice, self-hosting at that scale implies substantial multi-node GPU fleets and real serving expertise, so smaller organizations will likely consume it through hosted or partner deployments, while governments and regulated firms prioritize sovereignty. A likely trade-off of open weights is that guardrails can be modified or removed by whoever hosts the model, which is central to Mistral's pitch but also why it restricts the offensive-security variant. How the license terms handle redistribution is not detailed here.
Why it matters: Developers get an auditable, self-hostable frontier model with explicit cost controls via reasoning modes, which matters most to European enterprises and public bodies with data-residency constraints (inference). Security teams evaluating open-weight models for vulnerability triage and patching now have another benchmarked option, though the preview's caveat about attack uplift means defenders should expect guardrails to keep changing.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
The weights of Mistral Large 4 (Le Chonk) are now open. A 1T parameter multimodal model, and best open-weight model for coding, trained in Europe.