Xiaomi Unveils HySparse2 Architecture for MiMo-V3 to Boost AI Agent Efficiency
Xiaomi has disclosed HySparse2, the core architecture for its upcoming MiMo-V3 large language model, designed to optimize inference for long-context, multi-turn AI agent tasks. The new design introduces two-level Key-Value (KV) sharing and token-level sparse selection, enabling the prefill compute phase to exit after executing only the first half of the model's layers. By drastically reducing the computational and memory burdens of processing extended prompt histories, HySparse2 allows AI agents to handle tool outputs, code logs, and documents far more efficiently. At a 1-million-token context length, HySparse2 cuts prefill computation by 5x and reduces KV cache memory consumption from 12 GB down to 2.7 GB compared to standard Hybrid SWA architectures. HySparse2 splits the network into a Self-Decoder and a Cross-Decoder, leveraging KV Bridging to generate all required KV caches for the model's second half directly from the front half's hidden states. Additionally, it transitions from block-level to fine-grained token-level selection, mandatory-enforcing a 128-token local sliding window alongside 1,024 globally selected tokens.
## BACKGROUND
Large language models rely on attention mechanisms to calculate relationships between input tokens, generating Key and Value (KV) tensors stored in a KV cache for sequential token generation. In long-context tasks, traditional full attention requires quadratic computation and immense GPU memory to store the KV cache, making multi-turn AI agents expensive to run. Hybrid designs such as Sliding Window Attention (SWA) and sparse attention alleviate these bottlenecks by limiting full context calculations to select layers or tokens.