In short
- Mistral Large 4 (“Le Chonk”) is a model of about one trillion parameters with open weights, built for coding and security.
- Through the API it costs $0.70 per million input tokens and $2.10 output (Mistral's figures).
- The licence allows commercial use, but companies selling access to the model as a service need a separate licence.
In short: on 29 October 2026 Mistral AI released open weights for Mistral Large 4, nicknamed “Le Chonk”. It is a model of about one trillion parameters built for coding and security tasks. Mistral says it is the best open model for defending against attacks.
What changed
- The model combines a standard and a reasoning mode and accepts images; it has three reasoning levels (low, medium, high).
- The weights are on Hugging Face, and the model is available in Mistral Studio, through the API, OpenRouter, Microsoft Foundry and NVIDIA services.
- API pricing according to Mistral: $0.70 per million input tokens, $0.07 for cached input and $2.10 per million output tokens.
- The licence is a modified MIT: you may use, modify, fine-tune and sell, but a company offering the model as a service needs a separate commercial licence.
Results reported by Mistral
- Terminal-Bench 4.0: 41% (28% in the preview), Artificial Analysis Coding Agent Index: 48%.
- Security: 51% on the Cyber Index; 97.1% attack resistance on the B3 Agent Security test.
- In blind human preference tests participants preferred it over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash; only Claude Opus 5.5 ranked higher.
How to read it
These are the producer’s numbers, and the security vendors’ quotes on the page are partner statements, not independent tests. Mistral itself says that after the preview it saw “a potential for uplift” to real attacks and is changing its guardrails; a dedicated cybersecurity version is only for vetted partners. I write about cheaper models in the open models hub and about costs in the model pricing hub.