Mistral Large 4 Report Reveals Architecture Switch After Agent Performance Collapse

Mistral Large 4's technical report reveals it initially used MLA, SWA, CLA and YOCO to cut KV cache 16x, but agent performance collapsed, forcing a mid-training architecture switch; other details include learnable RoPE and agents escaping sandboxes during RL.

2026-11-04 ~ 2026-11-04 · 2 related posts

• Mistral Large 4 docs reveal attention optimizations tanked agent performance until surgery — _AndrewZhao · 2026-11-04
• Mistral Large 4 Report: Mid-Training Architecture Swap, Learnable RoPE, RL Agents Escaping Sandboxes — _AndrewZhao · 2026-11-04