Xiaomi's grand ambitions in large models have finally unveiled all their secrets in the spring of 2026.

On March 19, Xiaomi officially released three self-developed large models: MiMo-V2-Pro, MiMo-V2-Omni, and MiMo-V2-TTS. This is not just a routine technological upgrade but also a milestone for Xiaomi's full commitment to the "Agent era."

To demonstrate its determination to focus on AI, Xiaomi's founder clearly stated on a social platform that Xiaomi's R&D and capital investment in the AI field this year will exceed 16 billion yuan. He revealed that the trillion-parameter large model MiMo-V2-Pro ranks eighth globally in the comprehensive intelligence ranking of Artificial Analysis, and is fifth in brand ranking worldwide.

The three models each have their own roles, building a complete Agent stack:

Flagship base model MiMo-V2-Pro: Specifically designed for high-intensity Agent scenarios. Its total parameter count exceeds 1 trillion (1T), using a hybrid attention mechanism (Hybrid Attention), and achieves a balance between efficiency and capacity with 42B activated parameters. It supports an ultra-long context of 1 million Tokens, focusing on complex logical reasoning and tool calling.

Omni-modal base model MiMo-V2-Omni: Achieves native integration of text, vision, and audio. It connects the entire process from sensory understanding to action execution, making it key for Agents to perceive the physical world.

Speech large model MiMo-V2-TTS: Gives Agents the ability to express with "warmth," supporting fine-grained emotional control, making machine interactions more human-like.

In terms of commercialization, Xiaomi has adopted an extremely aggressive pricing strategy. The cost for input within 256K context is only $1 per million Tokens, significantly lower than competitors at the same level. Currently, both the Pro version and the Omni version have officially opened API services.

Notably, the mastermind behind this elite AI team is known as the "AI prodigy." The mysterious model "Hunter Alpha" that caused a stir in the developer community is actually the internal test version of MiMo-V2-Pro.