Xiaomi pushed its MiMo-V2.5-Pro model into public beta on Wednesday, positioning the release as a flagship for long-horizon agentic work. The announcement post frames V2.5-Pro as a step up from March's MiMo-V2-Pro in software engineering and multi-turn tool use.
The company's headline demos lean hard on coding endurance. V2.5-Pro built a complete SysY compiler in Rust from scratch in 4.3 hours across 672 tool calls, scoring 233 out of 233 on a Peking University course test suite. A separate run produced a working desktop video editor at 8,192 lines of code over 1,868 tool calls and 11.5 hours of autonomous work. Xiaomi also points to an analog-chip design task (an FVF-LDO regulator in TSMC 180nm) that it says the model closed in about an hour, with Claude Code acting as the harness.
The token-efficiency claim is the real pitch to developers. On Xiaomi's own ClawEval numbers, V2.5-Pro hits 64% Pass^3 using roughly 70,000 tokens per trajectory, which the company says is 40 to 60% fewer than Claude Opus 4.6, Gemini 3.1 Pro, and GPT-5.4 at similar scores. All benchmarks are self-reported.
Access during beta runs through the API and Xiaomi's web studio. OpenRouter lists pricing at $1 per million input tokens and $3 per million output, with a 1-million-token context window. Xiaomi says it plans to open-source the full MiMo-V2.5 series soon, though no date has been committed.
Bottom Line
MiMo-V2.5-Pro is live in API-only beta at $1 per million input tokens and $3 per million output, with weights promised but not dated.
Quick Facts
- 4.3 hours to build a complete SysY Rust compiler (672 tool calls, 233/233 tests, company-reported)
- 8,192 lines of code for the video editor demo, across 1,868 tool calls over 11.5 hours
- 64% Pass^3 on ClawEval at ~70K tokens per trajectory (Xiaomi self-reported)
- 40 to 60% fewer tokens than Opus 4.6, Gemini 3.1 Pro, and GPT-5.4 on the company's own benchmarks
- 1,048,576-token context window, priced at $1/M input and $3/M output on OpenRouter
