One-sentence summary
On October 29, 2026, France's Mistral AI released the weights of its flagship Mistral Large 4 (codenamed Le Chonk), a natively multimodal mixture-of-experts model with about 1.05 trillion total parameters and roughly 52 billion active per token, supporting a 1-million-token context window under a modified MIT open-weight license.
Event Overview
- Time: Weights released October 29, 2026 local time; the model was first publicly previewed on October 6.
- Company: French AI company Mistral AI.
- Event: The Mistral Large 4 (ML4, codename Le Chonk) checkpoint is downloadable from Hugging Face for self-hosting, and is served via the Mistral API, OpenRouter, NVIDIA NIM, and Microsoft Foundry.
Key Information
| Item | Detail |
|---|---|
| Total parameters | About 1.05 trillion (1T) |
| Active parameters | About 52B per token (49B excluding embeddings and output layers) |
| Architecture | Granular MoE: 128 routed experts, 4 active per token, plus 1 shared expert |
| Context window | 1 million tokens |
| Modality | Native multimodal: text + image input, text output; 1.6B vision encoder |
| Training | From scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters |
| License | Mistral Ecosystem Open-Weight License (modified MIT); MaaS providers need a separate commercial license |
| API pricing | Input $0.70/M tokens, cached input $0.07/M, output $2.10/M |
Background and Impact
- Positioning: Mistral calls Large 4 the strongest western open-weight model for coding and defensive cybersecurity; it scores 26 on the Artificial Analysis Intelligence Index v5, the highest among western open weights, while the company notes leading Chinese open-weight models score higher on the aggregate index.
- Guardrails: After monitoring live traffic post-preview, Mistral found the model refuses black-box attacks but sees potential for real misuse and is revising guardrails; average refusal rate on malicious cyber prompts is 96.6%.