Neural Brew

Models & labs

Mistral Releases Weights for Mistral Large 4 (“Le Chonk”), a One-Trillion-Parameter Open Model

Mistral AI has made the weights available for an upgraded version of Mistral Large 4, dubbed “Le Chonk”: a one-trillion-parameter multimodal mixture-of-experts model trained in Europe and presented as the best open model for code. The license is permissive, except for companies reselling access to the model as a service.

AI-generated portrait of LiorAI-generated portrait of Tomo

By and

Published · 4 min read

Abstract generated illustration for the Models & labs section

Mistral AI released the weights for Mistral Large 4, its flagship model nicknamed “Le Chonk,” this Thursday, October 29, at 3:21 p.m. Paris time. An upgraded version of the model that launched in public preview earlier this month, it is now available for download on Hugging Face. Mistral bills it as the top open-weight model for code and highlights its defensive cybersecurity capabilities.

What was released

According to the model card, Mistral Large 4 is a multimodal mixture-of-experts model totaling approximately 1 trillion parameters. For each token, only 49 billion parameters are activated (52 billion when factoring in input and output layers): the architecture selects 4 routed experts out of 128, along with 1 shared expert. It accepts text and image inputs, outputs text, and handles a 1-million-token context window across more than 160 languages, covering all official European Union languages.

Mistral underscores the model’s European roots: it was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in the developer’s own data centers across Europe. Compared to the preview release, the company states it added three weeks of reinforcement learning, yielding improvements concentrated in code and cybersecurity, as well as more token-efficient reasoning. The model provides three reasoning modes (low, medium, and high, with high set as default) to help users balance cost and precision.

Alongside the full-precision weights, Mistral published a lighter NVFP4 quantized variant designed for Blackwell GPUs. The model remains a massive workload: deployment configurations detailed in the model card target an 8-GPU node.

Figures claimed by Mistral

All metrics below were published by Mistral and have not been independently verified by Neural Brew:

  • 19.7% on FrontierCode 2.0, a private benchmark by Cognition that gauges whether a codebase maintainer would accept an agent’s proposed pull request; Mistral claims top placement among open models evaluated.
  • 48% on Artificial Analysis’s Coding Agent Index, matching GLM-5.3 according to Mistral, while utilizing 47% fewer tokens per task on average.
  • 41% on Terminal-Bench 4.0, up 13 points over the preview release according to Mistral.
  • 51% on Artificial Analysis’s Cyber Index, which evaluates capabilities in discovering, reproducing, and patching software vulnerabilities, alongside an average refusal rate of 96.6% on malicious cyberattack prompts.
  • 26 on Artificial Analysis’s Intelligence Index v5. The model card acknowledges that leading Chinese open models score higher on this aggregate benchmark, positioning Large 4 as the highest-ranking Western open model.

In a blind evaluation by Surge cited by Mistral, coding outputs from Large 4 were preferred over those from GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, though trailing Claude Opus 5.5.

An open license, with an exception

The model is distributed under the Mistral Ecosystem Open-Weight License, a modified version of the MIT License. Downstream users are permitted to run, modify, fine-tune, and redistribute it, including for commercial purposes. The explicit restriction applies to entities operating a “model-as-a-service” business—meaning those reselling access to the model via an API or hosted endpoints—which must execute a separate commercial license with Mistral. Internal enterprise usage and embedding the model within a specific application feature are exempt from this requirement.

Where to use it

Mistral lists the model on Mistral Studio, its API (identifier mistral-large-4), OpenRouter, its coding assistant Vibe Code, NVIDIA (hosted endpoints and NIM containers), and Microsoft Foundry, presented as the first cloud provider to offer it. Weights can be served using vLLM or SGLang. Just one day prior, Mistral rolled out an API compatible with OpenAI’s Responses format, as detailed in our October 28 article.

Cybersecurity focus, under scrutiny

Mistral had the model evaluated by security vendors including CrowdStrike, Cloudflare, Zscaler, and Rubrik, and also cites ENISA, the European Union Agency for Cybersecurity. The developer acknowledges an inherent risk: tracking real-world telemetry since the preview revealed “dual-use potential for real attacks,” even as the model rejected attacks against live infrastructure, leading the team to tighten its guardrails accordingly. A specialized offensive variant dedicated to penetration testing remains restricted to vetted partners, such as Orange Cyberdefense, under formal contractual agreements.

Our take

This marks Mistral’s most significant launch in months: a one-trillion-parameter model, fully trained in Europe and available for download, provides European corporations and government agencies with a credible self-hosted pathway—an offering Mistral markets as “sovereign AI.” While Large 4 does not eclipse top-tier Chinese open models on aggregate intelligence benchmarks, it squarely targets code and defensive cybersecurity, pairing this focus with token-efficiency gains critical for agentic workflows. The licensing structure is noteworthy: broadly permissive for internal deployments or embedded software features, it cuts off hosting providers looking to monetize model access without an agreement, insulating the developer’s own API business. Finally, the model’s physical footprint confines self-hosting to teams operating multi-GPU high-end clusters. Mistral’s performance claims still await validation through independent benchmarks.

Sources

  1. Mistral AI (X) — Le Chonk weights release announcement Primary source
  2. Mistral AI — Mistral Large 4 (Le Chonk) is now open weight Primary source
  3. Hugging Face — mistralai/Mistral-Large-4-1T-A52B model card Primary source
  4. Hugging Face — Mistral Ecosystem Open-Weight License text Primary source

A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.