Mistral Large 4 Weights Explained: Self-Hosting Le Chonk (2026)

This analysis was written by the aifreetool Editorial Team — a group of full-time AI-industry researchers and writers who verify every claim against primary sources. Last updated November 3, 2026. We keep no affiliate relationship with the companies covered here.

Quick answer: Mistral Large 4, codenamed Le Chonk, published its open weights on Hugging Face on October 29, 2026 — two days ahead of the promised date. It is a 1.05-trillion-parameter mixture-of-experts model with 52 billion active parameters and a 1M-token context window, released under a modified MIT license that lets most companies self-host it free of charge. What follows is self-hosting explained end to end: the specs, the license strings, the hardware math, and the cyber benchmark trade-off.

That second point is the real story. Five days before the weights dropped, Google shipped the closed Gemini 4 Carbon. Mistral went the opposite way and handed out a trillion parameters for download. For anyone tracking open models in our AI models directory, this release redraws the map of what you can run on your own servers.

What Mistral Actually Shipped on October 29

The official checkpoint is listed on Mistral's blog and hosted as mistralai/Mistral-Large-4-1T-A52B on Hugging Face. The architecture: 128 routed experts with 4 active per token plus 1 shared expert, a 1.6-billion-parameter vision encoder, and native multimodality. Mistral trained the model from scratch on roughly 3,800 NVIDIA Grace Blackwell GPUs inside its own European datacenters — a point the company repeats because it sells sovereignty as a feature, not a slogan.

Training data covered more than 160 languages, including every official EU language. Three reasoning modes ship out of the box: Low, Medium, and High, letting developers trade precision against token spend. Microsoft Foundry is the first third-party cloud to carry the model, and it is also served through OpenRouter and NVIDIA-hosted endpoints.

The 82% Cyber Score No Closed Model Will Match

Here is the number that should unsettle the safety-first crowd. On the Artificial Analysis Cyber Index test that asks a model to reproduce a real vulnerability in open-source software and then patch it, Mistral Large 4 scores 82% — the highest of any model tested. Claude Opus 5.5 and GPT-6 Astra score near zero on that same test, not because they cannot do it, but because they refuse. Defending software often starts with proving a flaw is real, and refusal-based guardrails block exactly that work.

Mistral is careful to frame this as defense. The model averages a 96.6% refusal rate on malicious cyber prompts across JailbreakBench, StrongREJECT, and AgentHarm — the highest the company has measured in any open model. On Cybench, a set of 40 security-competition exercises, it solves 93%. The company also posts a 51% score on the broader Artificial Analysis Cyber Index, which measures vulnerability discovery, reproduction, and patching without breaking working code. Whether you call that a guardrail or a loophole depends on which side of the firewall you sit.

Hardware Reality Check: Can You Actually Run It?

Most readers will not run this model at home, and the file sizes explain why. The table below summarizes the practical options as of this week.

A single RTX 5080 will not load even the smallest quantization. Community GGUF builds from Unsloth and ggml-org, plus vLLM and SGLang images, are already available, so the software path is paved — the hardware is the wall. For everyone else, API pricing starts at $0.70 per million input tokens, $0.07 for cached input, and $2.10 per million output tokens, which undercuts most frontier closed models by a wide margin.

The License: Modified MIT with One String Attached

The Mistral Ecosystem Open-Weight License is a modified MIT license. You can download, modify, fine-tune, and commercialize freely. The single condition: businesses that resell the model itself as a hosted service need a separate commercial license from Mistral. Running it inside your own product is fine; becoming someone else's model provider is not. That is a narrower restriction than most open-weight licenses in the enterprise space, but note that it is not Apache 2.0 like Mistral Large 3.

There is also a governance gap worth naming. Mistral's VP of Science, Pierre Stock, told CSO that guardrail updates will apply to the hosted API and future releases — not to checkpoints already downloaded. Once a trillion parameters leave the building, patches stop following them.

Key Takeaways

• Open weights went live October 29, 2026, beating Mistral's own October 31 deadline by two days.
• Specs: 1.05T total parameters, 49–52B active per token, 1M context, 128 experts with 4 routed plus 1 shared.
• 82% on vulnerability reproduction and patching — the top score of any model, closed or open.
• 96.6% refusal rate on malicious cyber prompts, so the defense-first framing has numbers behind it.
• License is modified MIT: free commercial use, except reselling it as a hosted model service.
• Smallest local build needs ~233 GB and a 256 GB RAM machine; most teams will use the $0.70/$2.10 API.

Frequently Asked Questions

Is Mistral Large 4 really open source? Not in the OSI sense. The weights are downloadable under a modified MIT license that permits almost everything, but hosted-service resale requires a separate commercial license, so it is best described as open-weight rather than open source.

How does it compare to GLM-5.3 or Kimi K3? On the Artificial Analysis Coding Agent Index it scores 48%, effectively tied with GLM-5.3 while using 47% fewer tokens per task. In Surge Code blind pairwise evaluations, human raters preferred it over GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash; only Claude Opus 5.5 ranked higher.

Can I run Mistral Large 4 on a gaming PC? No. Even the 1-bit GGUF quantization is about 233 GB and needs roughly 256 GB of RAM or unified memory. Practical self-hosting starts with a multi-GPU server, and the API is the sane default otherwise.

What does the cybersecurity 82% mean in practice? On one Artificial Analysis test, the model reproduced a real vulnerability in open-source software and then wrote a working patch, doing so on 82% of attempts — the highest score recorded, ahead of closed models that refuse the task outright.

My Take / The Bottom Line

Mistral Large 4 is the most consequential open-weight release of 2026, and the reason is not the trillion-parameter headline — it is the cybersecurity positioning. Mistral found the one workload where refusal-happy closed models score near zero and built a model that scores 82% while still refusing 96.6% of clearly malicious prompts. That is a coherent safety argument, and it doubles as a sales pitch for European buyers who want auditable security automation under their own roof. The losers here are closed-model vendors selling to security teams, and open-weight rivals outside China watching their lead evaporate. My advice: if your team does defensive security work, pilot the API this month and budget for a self-hosted deployment next quarter — but read the license before you plan anything that smells like reselling it.