IA Generativa
MiMo-v2.5-Pro: Xiaomi’s 1T Model Cuts AI Costs by 60%
Dr. Adrian Vale
LLM, AI Agents & AI Infrastructure Specialist
4 min read
Listen to Article
AI automated narration
LLM, AI Agents & AI Infrastructure Specialist
Xiaomi's MiMo-v2.5-Pro, a 1-trillion-parameter AI model, offers multimodal capabilities at a cost of $3 per million output tokens. With open-source availability and 1000 tokens/second generation speed, it challenges competitors like OpenAI with 60% cost savings and optimized performance.
Xiaomi has announced the MiMo-v2.5-Pro, a 1-trillion-parameter AI language model designed to bring cutting-edge capabilities to developers and enterprises at a fraction of the cost of its competitors. The model combines affordability, performance, and accessibility, marking Xiaomi’s bold entry into the global AI landscape.
The MiMo-v2.5-Pro is based on Mixture of Experts (MoE) architecture, which activates only relevant parameters for each task. This approach enhances efficiency without compromising performance.
In testing, MiMo-v2.5-Pro has shown competitive results against leading models:
A standout feature is its pricing:
This represents a 60% reduction in output token costs compared to GPT-5.4, which charges $6–$10 per million tokens. The MoE architecture plays a pivotal role in this cost optimization.
MiMo-v2.5-Pro supports text, image, audio, and video modalities, enabling advanced applications across diverse industries:
Xiaomi’s strategic move disrupts the AI market in several ways:
While the MiMo-v2.5-Pro shows immense potential, the following factors will shape its trajectory:
To drive adoption, Xiaomi has launched a 100 trillion token incentive program, aimed at encouraging developers to integrate and experiment with the model.
Xiaomi’s MiMo-v2.5-Pro stands as a formidable entrant in the AI landscape, combining massive scale, low costs, and multimodal capabilities. Its open-source nature and affordability could accelerate the democratization of AI and force incumbents to rethink their strategies. The next 6–12 months will be critical for gauging its adoption and long-term impact.
MiMo-v2.5-Pro costs $1 per million input tokens and $3 per million output tokens, which is 60% cheaper than competitors like GPT-5.4.
It features 1 trillion parameters, a 1-million-token context window, 1000 tokens/second speed, and supports multimodal inputs (text, image, audio, video).
Yes, MiMo-v2.5-Pro is fully open-source under the MIT license, with weights and tokenizers available on Hugging Face.
💡 Dica Pro: Leverage MiMo-v2.5-Pro’s hybrid attention mechanism (Sliding Window Attention and Global Attention) to optimize token generation speed without requiring specialized hardware.
