Overview
Mistral Large 4 (“Le Chonk”) is a 1.05T-parameter natively multimodal open-weight MoE model with 52B active parameters and 1.6B vision encoder. Preview API live October 2026; weights release end of October.
Specifications
- Training: 3,800 NVIDIA Grace Blackwell GPUs in European datacenters (~10MW)
- Context: 1M tokens
- API pricing: 2.09/M output (preview)
- Hugging Face: mistralai/Mistral-Large-4.0-1T05-A52B (expected Oct 27–31)
Claims (vendor-reported)
- Cybersecurity: 82% vulnerability reproduction; closed models refuse security tasks
- Agentic coding, finance/legal benchmarks competitive with strongest open models
- Visual grounding surpasses some frontier closed models
Warning
All benchmark claims require independent verification. Weights not yet publicly released.
Related
- mistral
- open-weight-models
- mixture-of-experts
- european-sovereign-ai
- ai-sovereignty
- ai-cybersecurity-testing
- multimodal-ai