Mistral open-weights Large 4: ~1T-param MoE with 1M context, free until you host it
AIFlow_ML · x · 2026-11-01
Mistral released the weights of Mistral Large 4 on Oct 29, the model teased as "Le Chonk".
- Architecture: MoE with 1T total parameters, 128 routed experts (4 active per token) plus one shared expert; 49B active parameters per token (52B including embeddings and output layers)
- Capabilities: natively multimodal, 1M-token context window
- License: open-weight and usable, but hosting it as a paid API is not allowed
- Context: an API preview shipped earlier in October; this checkpoint adds further reinforcement learning focused on coding and defensive cybersecurity
More from Models
- Rumor: OpenAI may drop Fable and GPT-6.5 as soon as Tuesday or Wednesday — iruletheworldmo · 2026-11-01
- ChatGPT Pro subscriber says the promised $100/month API credits are impossible to claim — Elerein · 2026-11-01
- Leak: 'Mythos 5.5' spotted in HTML source of Anthropic's official website — kimmonismus · 2026-11-01
- ChatGPT for Clinicians Finds 6 Significant Errors in Top Journal Paper — BraydonDymm · 2026-11-01
- Mistral License in Short: Free to Use and Ship, Paid API Hosting Needs a Deal — AIFlow_ML · 2026-11-01
- Supposed Anthropic 'Mythos-class' model leaks on HuggingFace — unverified — iamthe0ther0ne · 2026-11-01