MiMo-v2.5-Pro
6 hours ago
- #AI Model
- #Tech Innovation
- #Xiaomi
- Xiaomi released MiMo-V2.5-Pro in public beta, the most capable model to date with significant improvements in agentic capabilities, complex software engineering, and long-horizon tasks.
- MiMo-V2.5-Pro demonstrated strong long-horizon coherence, handling tasks over a thousand tool calls and improving instruction following in agentic scenarios, with enhanced context adherence.
- Key achievements include building a SysY compiler in Rust in 4.3 hours with perfect scores, creating a full-featured video editor in 11.5 hours, and designing an optimized analog circuit (FVF-LDO) within an hour.
- The model exhibited 'harness awareness', effectively using its environment and managing memory for structured, self-correcting workflows.
- Xiaomi enhanced coding intelligence through scaled compute, improving real-world coding scenarios and integration with frameworks like Claude Code and OpenCode.
- MiMo-V2.5-Pro achieved high efficiency, scoring 64% on ClawEval with about 70K tokens per trajectory, using 40–60% fewer tokens than competitors like Claude Opus 4.6 and GPT-5.4.
- Token Plan updates include resetting used Credit balances for users before April 21, 14:00 UTC, alongside upgraded inference infrastructure.
- Xiaomi announced plans to officially release and open-source the MiMo-V2.5 series soon.