MiMo-V3 is getting a new architecture. The core of it, HySparse2, is out today.
This is a dev post classified by Jev as Other (a launch), kept by the Dev Radar because it carries real work, not commentary.
MiMo-V3 is getting a new architecture. The core of it, HySparse2, is out today. Less prefill, a smaller KV cache, better long-context retrieval—and we got all three at once. Compared with MiMo-V2.6's Hybrid SWA architecture: • 5.02× lower prefill FLOPs at 1M tokens • 4.5× smaller KV cache at 1M tokens • Better MRCRv2 and RULER-v2 scores, plus lower AgentPPL and LongPPL Why build a new architecture? Agentic inference is a very different workload. Each round, a short action can return a long observation that needs to be prefilled, while the context keeps growing. That puts prefill cost, KV-ca
Posted by Fuli Luo (84k followers) 1 h ago · 936 likes · 28.8k views · view the original post on X. Kept by the Dev Radar as Other.
More dev work like this
- Eikos, um modelo de decisão open source (MIT), alternativa ao Jev e ao Laya. — @0xCVYH
- The name HySparse2 really makes me think of Hy (Hunyuan). I even catch myself reading it… — @sheriyuo
- Not bad — @zephyr_z9
- Xiaomi is skipping a generation. V2.6 still uses MiMo Hybrid-SWA, a very 2025 design;… — @teortaxesTex
- This is awesome! — @ItsmeAjayKV
- China Mobile just open-sourced a bridge between Physical AI models and humanoid robots. — @CyberRobooo
- China just dropped a Claude Opus 5 level model that runs locally. — @thesupermannx
- I’m building something awesome. 👀 — @LeSiOO
Every post is read and classified by Jev (TypeSafe): what it is, which market it belongs to, and whether the link is a real tool. 22.6k posts from 5k X accounts over the last 21 days, 2.6k tools, 12 markets. Collected every 5 minutes, fully re-ranked every hour — last update 2026-09-23 15:21 UTC. Full method.