Mistral Large 4 tech report turned into visual guide: 1T MoE, 52B active
_AndrewZhao · x · 2026-11-07
The unusually candid Mistral Large 4 tech report has been converted into a scrollable visual guide. Key facts: 1T-parameter MoE with 52B active per token, built for reasoning and long-horizon agentic work; an attention design chosen for cheap serving hurt agentic performance and was surgically fixed mid-training; agent-generated training tasks, async RL with documented reward hacks; scores include Terminal-Bench 4.0 at 41%; serving fits 606 GB on 4×GB300.
More from Models
- 4B Nemotron 3 Nano claimed to beat frontier models on task-specific workflows — eliano · 2026-11-07
- Thinking Machines' Inkling 2 Small hits 81.7% on ARC-AGI-2, full weights coming next week — ziqiao_ma · 2026-11-07
- LithosAI claims fastest Kimi K3.1 Pro inference at 4.7 min per Intelligence Index task — JiaZhihao · 2026-11-07
- Applied Compute shares recipe for production RL with continual learning — ypatil125 · 2026-11-07
- Qwen3.8 27B writes 1,030-word replies, 6x longer than GPT-6 Sol's 174 — ArtificialAnlys · 2026-11-07
- China Telecom's Xing4.0-29B-A4B open-sourced: trained on Ascend NPUs, 75 on SWE-bench Verified — jacek2023 · 2026-11-07