Mistral Large

Editor's Review

Mistral Large 4 — released October 6, 2026, codename "Le Chonk" — is Europe's flagship AI bet: a 1.05T-parameter multimodal MoE (49B active per token) with a native 1.6B vision encoder, 1M-token context window, 160+ languages, trained on 3,800 NVIDIA Grace Blackwell GPUs in EU datacenters. API pricing: $1.36/$4.18 per million input/output tokens, with half-price preview rates live now ($0.68 in / $2.09 out / $0.07 cached). Open weights are promised by end of October 2026 (VentureBeat reports October 27) — currently API-only. The honest benchmark read: Artificial Analysis Intelligence Index scores it 38 — behind Chinese open-weight rivals GLM-5.3 (45) and Kimi K3 (44), ahead only of DeepSeek V4 Pro (36). Its notable strength is cybersecurity benchmarks, where closed models often refuse tasks. The license for the coming weights is unannounced (Large 3 was Apache 2.0; reports say Large 4 gets a custom Mistral license — verify before planning commercial self-hosting), and served context is 512K in preview despite the 1M documentation claim. Who should care? EU-sovereignty buyers and developers betting on the open-weights roadmap — if the October 27 weight drop lands under a permissive license, it's a major event. For API buyers today, the pitch is price: it undercuts closed flagships (GPT-5.6 Sol $4/$20, Claude Opus 5 $5/$25) by 3–6× with far better cache-read economics for agent loops. Versus DeepSeek V4 Pro (1.6T params, MIT weights already on Hugging Face): DeepSeek's weights are live and licensing is clear; Mistral's are still a promise. Don't build on promises — evaluate the API today, revisit after the weight release.