MiMo-V2.6 exhibits tool-call repetition where the model issues identical or nearly identical tool calls repeatedly without meaningful progress, affecting user experience across multiple agent platforms. Internal evaluations revealed repetition rates exceeding 0.05%, with the issue emerging during reinforcement learning training as a reward blind spot in optimizing for correctness.
MiMo-V3 introduces HySparse2, a new architecture designed for agentic inference that reduces prefill FLOPs by 5× and KV cache size by 4.5× compared to MiMo-V2.6 while improving long-context retrieval. HySparse2 uses two levels of KV sharing—KV Bridging and KV Reuse—along with token-level selection and a unified cache for local and global tokens to optimize the workload of processing short actions followed by long observations.