Mistral Large 4 lands: 1.05T-param MoE, best Western open weights but only mid-pack
maier_ak · x · 2026-11-09
Andreas Maier's deep dive on Mistral Large 4, the model born from a fan hoax:
- Origin: In June, Reddit/X users invented a fake "Le Chaton Fat" model (30T params, 1,000 meows/sec) with fake benchmarks; CEO Arthur Mensch played along. On Oct 6 Mistral officially named its new model "le Chonk," and weights dropped on Hugging Face on Oct 29 as promised.
- Specs: MoE with 1.05T total params, 49B active per token; routes through 4 of 128 experts plus a shared one; text+image input, 1M context, three reasoning-effort levels. Trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters on 160+ languages. API: $0.70/M input, $2.10/M output tokens.
- Benchmarks: 26 on the Artificial Analysis Intelligence Index (10th of 23) — best Western open-weight model but behind leading Chinese models; Coding Agent Index 48%, tied with GLM-5.3.
- Caveats: its record 82% on a cyber reproduce-and-patch test partly reflects closed models refusing the task; the license is "open with an asterisk" — MaaS providers need a separate deal.
More from Models
- 18 of Top 20 Open Models Use MoE; HuggingFace Spent 800K GPU-Hours on a Free Handbook — FinanceYF5 · 2026-11-09
- ValsAI reports critical safety flaw in frontier lab models, may delay releases — koltregaskes · 2026-11-09
- Tesla Optimus engineer builds GROXIE, a countertop robot assistant powered by Grok — Scobleizer · 2026-11-09
- ChatGPT-User crawls drop 81% as OpenAI shifts to its own index — gaganghotra_ · 2026-11-09
- LLMs show "insane regression" in intelligence when discussing cell therapy, says blogger — owl_posting · 2026-11-09
- Simon Willison: Claude Opus 5.5 Ports Three.js 3D Scene to Vanilla JS + WebGL — Simon Willison · 2026-11-09