Mistral Large 4 (Le Chonk) becomes a 1-trillion-parameter open model

Mistral AI released the weights of Large 4, codenamed Le Chonk, a multimodal model with 1 trillion parameters that leads open code benchmarks, with a high refusal rate for malicious security requests. For those looking to escape closed APIs, the announcement comes with important caveats about infrastructure and licensing.

Mistral AI released the weights of Mistral Large 4, internal codename Le Chonk, a natively multimodal model with 1 trillion parameters. The release opens up a model that had entered public preview in early October 2026, and which the company presents as the most capable open-weight model in the world for coding, with defensive security as one of its stated pillars, according to the official announcement.

Large 4 was born big, natively multimodal, and, according to Mistral, trained from scratch on European infrastructure, with a clear lineage from conception to launch.

What changed since the preview

Mistral says that a few additional weeks of reinforcement learning since the preview produced significant gains in coding and defensive cybersecurity, with no signs of saturation. The company also highlights an improvement in reasoning efficiency, which, it says, reduces inference cost without sacrificing performance. This is the kind of detail that matters for those who will host the model and pay the GPU bill per token processed.

Where Large 4 leads among open models

In Surge Code's blind, paired evaluations, with human raters comparing code responses side by side, Large 4 was preferred over GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, trailing only Claude Opus 5.5 among the models tested. Mistral also reports 84.5% on BrowseComp (autonomous web browsing), up from an SFT checkpoint that scored 75%, surpassing GLM-5.3 and DeepSeek Flash 4.1 1017. On FinWorkBench financial tasks, the model reached 73.3%, trailing only Kimi K3 among open models.

Defense and attack: what changes in offensive security

On Artificial Analysis's Cyber Index, which measures vulnerability discovery, reproduction, and remediation without breaking existing functionality, Large 4 scores 51%, without Mistral claiming leadership on this specific benchmark. Mistral says the model retains this capability without unnecessary refusals on legitimate defensive tasks, with an average refusal rate of 96.6% on malicious requests (measured across the JailbreakBench, StrongREJECT, and AgentHarm benchmarks), the highest among the open models the company measures.

Several security companies tested the model before launch. Cloudflare's statement sums up the central argument behind the "sovereign AI" category that Mistral is trying to establish:

When threats move at machine speed, human defense cannot keep pace. Worse, a defense you don't control isn't defense—it's exposure. Security teams perform best when they take the steering wheel. Mistral's ML4 lets defenders audit the architecture, deploy on their own infrastructure, and keep full sovereignty. Resilience starts with autonomy, and open-weight models help give organizations the control to lead their own defense.

Stephanie Cohen, Chief Strategy Officer at Cloudflare

Mistral itself admits a delicate point: while monitoring preview traffic, it observed the model refusing black-box attacks against real targets, but identified potential uplift for real-world attacks. The company says it is adjusting the guardrails and will continue updating them in the coming months. It is also giving partners such as Orange Cyberdefense early access to a specialized variant for pentesting and red-teaming, under rigorous vetting and restricted to authorized systems.

Sovereignty first, and that comes with an infrastructure price

The word that appears most often in Mistral's announcement isn't performance, it's sovereignty. The company is selling Large 4 to organizations that want to inspect, customize, and run the model on their own infrastructure, without depending on Mistral's API. Banks such as La Banque Postale, ABN AMRO, and Crédit Agricole, along with the European space agency (ESA Φ-lab), are cited as adopters already using or evaluating the model under this scheme.

In practice, this isn't the same as "running locally" in the sense an individual developer usually means by the term. A model with 1 trillion parameters is, by its scale, much closer to the reality of a bank or carrier with its own GPU cluster than to someone downloading weights for a workstation. Mistral doesn't detail hardware requirements in the announcement, but the model's size alone already defines the real audience for this release: platform teams with infrastructure budgets, not individual developers testing at home.

For those in this audience, Mistral offers two products to support customization: Mistral Forge, for adapting the model to proprietary data and security workflows, and Mistral Studio, for building and governing the resulting agents. Well-defined cybersecurity tasks can also be distilled into smaller, more specific variants, according to the company.

Three reasoning modes to calibrate cost

Large 4 ships with three reasoning modes that developers choose based on the task:

• Low: simple tasks, replaces the old no-reasoning mode
• Medium: everyday tasks
• High (default): harder cases that require precision and full reasoning

This explicit control matters for anyone running the model in production: it means deciding, by route or by request type, when it's worth paying the extra cost of reasoning tokens and when a cheaper mode will do. Mistral also showed, across three demos, a vulnerability triage agent that reproduces flaws in a sandbox and iterates until the patch works, a complete satellite tracking app (Orbital Trace) built end to end by an agent with no human help, and a quantum chemistry solver written from scratch by the model itself.

What remains open

The announcement doesn't detail the exact licensing terms for commercial use of the released weights, nor does it include a pricing table for those who choose to run via API instead of self-hosting. There's also no specification of minimum hardware requirements, nor a direct comparison with closed models like GPT-4o or Gemini, something that, given Mistral's stated ambition to compete at the top of the category, leaves a gap for those who need to decide between giving up a managed API and taking on the operation of a model this size.