Xiaomi Previews MiMo-V3 with HySparse2 Architecture, Cutting Prefill Compute 5x

Xiaomi MiMo lead Luo Fuli previewed MiMo-V3's core architecture component HySparse2, with the technical paper published simultaneously on arXiv (2609.26368), only two days after MiMo-V2.6's release. HySparse2 targets multi-turn agent workloads where each turn's short action returns long observations and context keeps growing, making prefill cost, KV cache and retrieval accuracy simultaneous bottlenecks; it claims 5x lower prefill compute for million-token inputs and KV cache reduced to 1/4.5.

Confirmed

Officially disclosed by Luo Fuli; paper public on arXiv as 2609.26368
Multi-layer redesign for long agent tasks; QbitAI mentions "prefill early exit" as one component, splitting the model into a front self-attention section among others (details not fully elaborated in posts)
Claimed metrics: 5x prefill compute reduction, KV cache at 1/4.5
MiMo-V2.6 still uses MiMo Hybrid-SWA

Unconfirmed

Technical details relayed in posts are incomplete; specifics require the original paper
Observers (relayed by teortaxesTex) argue Xiaomi is "skipping generations": HySparse, like NSA and MoBA, is an engineering of published research ideas, allegedly mirroring DeepSeek's NSA-to-DSA path—observer opinion, not official

Why it matters

Revealing the next-gen architecture two days after the previous release is unusually aggressive, signaling heavy investment in agent inference efficiency
Prefill compute and KV cache are core deployment cost bottlenecks for multi-turn agents; if the numbers hold, long-context agent inference costs drop significantly

2026-09-23 ~ 2026-09-24 · 5 related posts

Primary sources

Xiaomi Previews MiMo-V3's HySparse2 Architecture, Cutting Million-Token Prefill Compute ~5x — 量子位 · 2026-09-24 12:42 (UTC+08:00)
Xiaomi Previews HySparse2 for MiMo-V3: KV Cache Cut to ~1/4.5 at 1M Context — 智东西 · 2026-09-24 10:11 (UTC+08:00)
Xiaomi's MiMo-V3 gets new HySparse2 architecture: 5x prefill FLOPs cut, 4.5x smaller KV cache — teortaxesTex · 2026-09-23 23:01 (UTC+08:00)

MiMo-V3 to get new architecture, core HySparse2 paper published on arXiv — Recoil42 · 2026-09-23
[source] Xiaomi's MiMo-V3 gets new HySparse2 architecture: 5x prefill FLOPs cut, 4.5x smaller KV cache — teortaxesTex · 2026-09-23
Xiaomi's MiMo sparsity upgrade seen as mirroring DeepSeek's NSA-to-DSA move — teortaxesTex · 2026-09-24
[source] Xiaomi Previews HySparse2 for MiMo-V3: KV Cache Cut to ~1/4.5 at 1M Context — 智东西 · 2026-09-24
[source] Xiaomi Previews MiMo-V3's HySparse2 Architecture, Cutting Million-Token Prefill Compute ~5x — 量子位 · 2026-09-24