Xiaomi Announces HySparse 2 Architecture Core for MiMo-V3 LLM
Fuli Luo, head of Xiaomi's MiMo LLM team, announced HySparse 2, the new core architecture for the upcoming MiMo-V3 model. Designed for long-context and agentic workloads, HySparse 2 reduces prefill compute (FLOPs) by 5.02x and KV cache memory footprint by 4.5x at a 1-million-token context length. As agentic workflows accumulate massive context histories through multi-turn tool interactions, prefill latency and memory limits severely constrain real-world deployment. By dramatically lowering resource demands for million-token contexts, HySparse 2 makes executing complex long-context reasoning and AI agent tasks significantly faster and cheaper. HySparse 2 builds on a YOCO-inspired decoder-decoder structure using KV Bridging and KV Reuse, enabling token-level sparse selection and allowing prefill computation to exit early once the self-decoder finishes. Benchmark results demonstrate higher accuracy scores on RULER-v2 and MRCRv2 alongside improved (lower) AgentPPL and LongPPL perplexity metrics.
## BACKGROUND
In Transformer-based large language models, the Key-Value (KV) cache stores past token states to speed up text generation, but its memory consumption grows linearly with prompt length. Architectures like YOCO ("You Only Cache Once") address this by using a split decoder design where cross-decoder layers reuse states from a self-decoder, saving KV memory and enabling early-exit prefilling.