Mistral has released the open weights of Large 4, its 1-trillion-parameter multimodal flagship, unofficially nicknamed "Le Chonk." First shown as a public preview in the middle of the month, the model is now downloadable and has been re-priced for API use. Mistral is positioning it as the strongest open-weight model for coding and defensive cybersecurity, and as a European sovereign-AI option that organizations can inspect, customize, and self-host.

A 1T-Parameter Open-Weight Flagship

Large 4 is a mixture-of-experts model with 1 trillion total parameters and 52 billion active parameters, natively multimodal with vision input. On Hugging Face it is published as mistralai/Mistral-Large-4-1T-A52B, and Mistral says it can be deployed on your own infrastructure with vLLM or SGLang.

According to Mistral, the release builds on roughly three weeks of additional reinforcement learning after the preview, driving gains in coding, defensive cybersecurity, and reasoning efficiency. The company stresses its sovereign-AI credentials: the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs inside Mistral's own datacenters in Europe, and its training data spanned more than 160 languages, including every official language of the European Union.

Coding: Top of the Open-Weight Benchmarks Mistral Cites

Mistral reports that the release has moved Large 4 to the front of open-weight coding. On Cognition's private FrontierCode 2.0 benchmark, which scores end-to-end code quality by whether a maintainer would merge the change, the company says Large 4 leads with 19.7%. On the Artificial Analysis Coding Agent Index it scores 48%, which Mistral describes as effectively tied with the leading open-weight model, GLM-5.3, while using 47% fewer tokens per task on average. On Terminal-Bench 4.0 it improved 13 percentage points to 41% over the preview, cutting output tokens by 30%.

To let developers trade off performance against cost, the model now ships with three reasoning modes: Low, Medium, and High (default). In blind pairwise human evaluations run by Surge Code, Mistral reports that Large 4 was preferred over GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, ranking behind only Claude Opus 5.5 among the models tested. These benchmark figures are self-reported by Mistral; independent replication is pending.

Built for Defenders

Mistral has been evaluating Large 4 with cybersecurity vendors on real defensive work: discovering vulnerabilities, reproducing them, and patching without breaking functionality. The company reports 51% on the Artificial Analysis Cyber Index, and says the model's guardrails produce an average refusal rate of 96.6% on malicious cyber requests, the highest it measures among open models.

Mistral also notes that its guardrails distinguish defensive from offensive tasks, and that it is continuing to modify them after observing, in live traffic, a potential for the model to be used to conduct real attacks. Enterprise design partners such as Orange Cyberdefense are getting preview access to a specialized variant aimed at penetration testing, vulnerability assessment, and red-teaming, with access subject to vetting and contractual safeguards and testing restricted to owned or explicitly authorized systems.

Enterprise Results and Availability

Mistral reports broad enterprise gains as well: 73.3% on FinWorkBench (Finch), which it says is second only to Kimi K3 among open models for spreadsheet and finance workflows; 84.5% on BrowseComp for web-browsing agents; and 97.1% resistance on the B3 Agent Security Benchmark, up from 93.3% at preview.

The model is released under the Mistral Ecosystem Open-Weight license, a modified MIT: free to use, modify, fine-tune, and commercialize, with one condition that businesses operating a Model-as-a-Service offering must obtain a separate commercial license from Mistral. Starting October 29, API pricing is $0.70 per million input tokens, $0.07 per million cached-input tokens, and $2.10 per million output tokens, down from the preview's $1.36 / $4.18. Large 4 is available in Mistral Studio, via the Mistral API and OpenRouter, as open weights on Hugging Face, through NVIDIA NIM and TensorRT-LLM, and via Microsoft Foundry, which Mistral describes as the first cloud to carry the model, with private Azure deployment options.

Why This Matters

Large 4 is one of the largest open-weight MoE models to date and lands in the middle of a crowded open frontier that includes GLM-5.3, Kimi K3, DeepSeek V4, and Qwen3.8. Its open weights, European training, and self-hosting path make it a direct competitor in the sovereign-AI category that Mistral is trying to define. Two caveats for readers: the headline coding and cybersecurity numbers are self-reported and await independent verification, and the MaaS license condition is a material constraint for any cloud provider that wants to resell the model.

Sources

  • 1T-parameter multimodal MoE (52B active), open-weight release October 29, FrontierCode 2.0, Artificial Analysis Coding Agent Index, Terminal-Bench 4.0, three reasoning modes, cyber-safety results, FinWorkBench/BrowseComp/B3 Agent Security, pricing, availability: Mistral Large 4 (Le Chonk) is now open weight
  • Trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European datacenters; 160+ languages: Introducing Mistral Large 4