Xiaomi opens public beta of MiMo-V2.5 models
Xiaomi has opened a public beta of its MiMo-V2.5 model series, saying its flagship mimo-v2.5-pro can match Claude Opus 4.6 and GPT-5.4 on long-running agentic and coding work.
Xiaomi announced a public beta of its MiMo-V2.5 series. The lineup has four parts: mimo-v2.5, mimo-v2.5-pro, a V2.5-TTS series and v2.5-asr. The page carries a "V2.6" headline, but its text describes only the V2.5 release.
Xiaomi calls mimo-v2.5-pro its most capable model so far. It says the model can run nearly a thousand tool calls in a single session. In one internal demo, the model built a SysY compiler in Rust, a project from a Peking University course. Xiaomi says the work took 4.3 hours and 672 tool calls and scored 233/233 on a hidden test set. In a second demo, the model built a web video editor: 8,192 lines of code, 1,868 tool calls and 11.5 hours of autonomous work. Xiaomi ran these demos itself, and no one outside the company has checked the results or the claims of parity with Opus 4.6 and GPT-5.4.
The base mimo-v2.5 model is natively multimodal, handling images, audio and video, with a context window of up to 1M tokens. Xiaomi says it beats the previous mimo-v2-pro on the Claw-Eval agent benchmark at roughly half the API cost. Xiaomi also reports competitive scores on VideoMME, CharXiv and MMMU-Pro. To reach the same ClawEval scores, the company says v2.5-pro uses 42% fewer tokens than Kimi K2.6, and v2.5 uses 50% fewer than Muse Spark.
Xiaomi also changed its Token Plan pricing:
- Credit rates: mimo-v2.5 costs 1 credit per token and v2.5-pro costs 2. The 4-credit rate is gone, and 256k and 1M context windows now cost the same.
- Night discount: credit use drops a further 20% from 00:00 to 08:00 Beijing time.
- Auto-renewal: a new monthly auto-renewal option gives a discount, but the details are cut off in the source.