March 19, early this morningMilletThe Mimo V2 series was published in a centralized seriesModelMiMo-V2-Pro, Agent Base Mimo-V2-Omni, and Mimo-V2-TTS。
MiMo-V2-Pro:
The total amount of the parameter exceeds 1T, the activated parameter is 42B, using the Hybrid Attention structure, the mixed ratio was raised from 5:1 to 7:1 in the previous generation, and support was provided to the 1M super-long context window
In the framework of intelligent bodies such as OpenClaw and Claude Code, complex workflow programming, long-range planning and precision tools can be done without manual intervention
In the context of Coding capacity, the inner rice engineer assesses that it is close to Claude Opus 4.6 and has enhanced system design and mission planning capabilities
API pricing is about 1/5,256K in the context of Claude peer model $1/ million token, output $3/ million token; 1M enter $2/ million token, output $6/ million token in the context。
MiMo-V2-Omni is the first full-modular model for the integration of perception and action at the base level of millimetres, integrating text, visual and voice input, supporting the 256K context, priced as input $0.4/ million token, output $2/ million token。
Officially, it is claimed that its combined audio understanding is going beyond Gemini 3 Pro and image is going beyond Claude Opus 4.6; in the practical application scenario, MiMo-V2-Omni can automate browser operations in conjunction with the OpenClaw framework, including cross-platform pricedown, short video production and distribution end-to-end tasks。
The MiMo-V2-TTS speech synthesis model supports multi-particle control of emotions from the whole style to the rest of the sentence, allows for the natural processing of signals in the form of markers, words, etc., and supports a variety of dialects, such as Northeast, Sichuan and Sichuan, as well as role-playing styles and sound synthesis。
All three models are connected to the Kyinsan WFS, MiMo-V2-Pro Synchronized Microphone Client Agent product miclaw and milli browser。