MiMo-v2.5-Pro-UltraSpeed: 1T model with 1000 tokens per second (mimo.xiaomi.com)

Xiaomi has launched the MiMo-V2.5-Pro-UltraSpeed, a groundbreaking AI model capable of decoding over 1000 tokens per second with a massive 1 trillion parameters. This release, developed in collaboration with TileRT, is available via a limited-time API at a promotional price, significantly enhancing generation speed—offering around ten times the output experience compared to its predecessor. This unprecedented speed opens up new possibilities for real-time AI applications, enabling scenarios like high-frequency trading, instant fraud detection, and even medical assistance where every second counts. The model achieves its remarkable performance through innovative techniques in model-system codesign, employing FP4 quantization to reduce memory demands and DFlash for efficient speculative decoding. Unlike traditional methods that require specialized hardware, MiMo and TileRT have optimized the model to run efficiently on commodity GPUs. This allows for parallel reasoning paths and greater productivity in coding tasks, fundamentally transforming how AI models can interact in time-sensitive environments. With this advancement, speed becomes more than just a metric—it's a key enabler for delivering intelligent, real-time solutions that can significantly impact various industries.