🚨重磅！ MiMo-V2.6 刚登顶开源，小米AI负责人罗福莉直接预告 MiMo-V3换心脏！
This is a dev post classified by Jev as Other (a launch), kept by the Dev Radar because it carries real work, not commentary.
🚨重磅！ MiMo-V2.6 刚登顶开源，小米AI负责人罗福莉直接预告 MiMo-V3换心脏！ MiMo-V3 将上全新架构 HySparse2。不是微调，是把 Agent 时代最贵的三件事一起砍掉： 🔹1M tokens Prefill 算力再降 5.02× 🔹KV Cache 缩小 4.5×（12.09GB → 2.69GB） 🔹长文检索反而更准：RULER-v2 / MRCRv2 全面反超 Hybrid SWA，AgentPPL、LongPPL 全程更低 为什么必须换架构？ Agent 不是聊天。短动作 → 超长 observation → 上下文无限涨。Prefill、显存、检索精度同时卡死。 HySparse2 怎么破： 1️⃣KV Bridging：YOCO 式双解码器，只桥接 Full Attention，cross-decoder 的 KV 全由 self-decoder 投影出来 2️⃣KV Reuse：稀疏层直接复用上一层 Full Attention 的 KV 和选中 token 3️⃣块级稀疏改 token 级，强制近期窗口，局部+全局共用一份 Cache 4️⃣Prefill 跑完 self-decoder 就能停，后半模型直接跳过 一句话：半个模型就能把百万级上下文的 KV 建完。 论文今天已出，V3 还没发，架构先落地。 查看论文：h
Posted by SuSu_酥酥👅 (16.7k followers) 1 h ago · 17 likes · 2.1k views · view the original post on X. Kept by the Dev Radar as Other.
More dev work like this
- CCTP is expanding beyond USDC. — @circle
- Run uncensored Bonsai 2 27B on 16 GB MacBook 💻 — @atomic_chat_hq
- Marking our third audio model release in < 30 days, today we're introducing two new… — @NewsFromGoogle
- Meet XYZ-Aquila-mini-APEX-I-MiniPlus-V2.1-GGUF: a quantized MoE model built for… — @HuggingModels
- 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
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 16:31 UTC. Full method.