- Xiaomi has released its flagship AI model, MiMo-V2-Pro, which boasts over 1 trillion parameters and is designed to handle high-intensity agent workflows.
- The previously leaked mysterious model, Hunter Alpha, has now been confirmed as an early version of the Xiaomi AI model.
Xiaomi officially launched its AI model, Xiaomi MiMo-V2-Pro, on Thursday, specifically designed for high-intensity, real-world agent workflows.
The model features over 1 trillion total parameters and includes 42 billion active parameters. Its underlying architecture inherits and significantly optimizes an innovative mixed attention mechanism, the tech giant announced today.
The new model improves inference efficiency and supports an ultra-long context length of up to 1 million tokens, which significantly enhances the system's ability to handle multi-step complex reasoning, the company said.
The mysterious model "Hunter Alpha," which was previously released anonymously on developer platforms and quickly topped the API call charts, has now been officially confirmed as an early version of this model.
On the Artificial Analysis global comprehensive intelligence leaderboard, MiMo-V2-Pro currently ranks eighth globally and second among Chinese models.
The model demonstrates outstanding end-to-end task completion capabilities in complex workflow orchestration and precise tool calling, with its overall user experience approaching leading industry standards.
Xiaomi has now officially opened the API service for the model, adopting a highly competitive tiered pricing strategy based on usage, which greatly lowers the barrier to accessing cutting-edge AI.
Within a 256K context window, its input and output costs are $1 and $3 per million tokens, respectively, while the input cost within a 1M context window is $2.
In addition to the Xiaomi MiMo-V2-Pro, Xiaomi has also launched the MiMo-V2-Omni and MiMo-V2-TTS.
The MiMo-V2-Omni is positioned as Xiaomi's multimodal foundation model, integrating multimodal understanding capabilities for images, videos, and audio with powerful agent capabilities.
MiMo-V2-TTS, on the other hand, is geared toward the era of voice-based agents, offering highly controllable, fine-grained style control, natural prosody reproduction, and singing capabilities.
This concentrated release indicates that Xiaomi is rapidly transitioning from the AI R&D investment phase to the comprehensive realization of actual commercial results, Goldman Sachs said in a research note on Thursday.
The investment bank maintained a "buy" rating on Xiaomi's stock and set a 12-month target price of HK$41 ($5.23), implying an upside potential of about 14% for the company's shares.
Xiaomi's founder, chairman, and CEO Lei Jun said on Weibo today that the company's actual progress in the AI field has exceeded expectations, and the total amount of AI R&D and capital investment this year will surpass 16 billion yuan ($2.3 billion).
The new AI models has been deeply integrated into the Kingsoft Office ecosystem and Xiaomi's smartphone operating system, aiming to reshape the end-user agent experience through cross-device collaboration.
Looking ahead, the core focus of Xiaomi's R&D team is to conquer high-complexity reasoning and long-term task planning, in order to systematically enhance the model's decision-making capabilities in unknown environments, the company said.
($1 = 7.8386 HKD, $1 = 6.8959 yuan)
