Xiaomi opens API for MiMo-V2-Omni agent model
Xiaomi has made MiMo-V2-Omni, a model that handles text, vision and speech and can carry out agent tasks, available through a low-cost public API, though its benchmark claims are self-reported and come without figures.
Xiaomi announced MiMo-V2-Omni, which it calls an "omni-modal foundation model for the agent era." The company says the model was built from the start as one architecture covering text, vision and speech. It also comes with built-in tool calling, function execution and GUI operation, and Xiaomi says it works with the major agent frameworks.
The page is titled "MiMo-V2.6 Series: 3 New Models Officially Released," but the text describes only MiMo-V2-Omni. It does not name the other models or say what "V2.6" refers to.
Xiaomi's performance claims are relative and come without numbers:
- Vision: better than Claude 4.6 Opus on multidisciplinary visual reasoning and chart analysis, and "closing the gap" with Gemini 3.
- Audio: better overall than Gemini 3 Pro, with support for continuous audio longer than 10 hours.
- Agent tasks: on par with Gemini 3 Pro on benchmarks that involve interacting with real digital environments, and "highly competitive" on text-only agent tasks.
The page does not name the benchmarks or give scores, model size or architecture details.
Xiaomi's demos include a browser agent that researched products on Xiaohongshu, compared prices on JD, haggled with a human customer-service agent and placed an order. In another, the model made and published a TikTok video, fixing a font-rendering error along the way. Xiaomi also says the model can produce Word, Excel, PDF and PowerPoint files that are close to final.
The API is live at platform.xiaomimimo.com. It costs $0.40 per million input tokens and $2 per million output tokens.