The generative AI industry is currently hitting a severe processing bottleneck, as long wait times restrict the real-world usefulness of autonomous coding agents. Xiaomi just shattered that limit. On Monday, the company officially released MiMo-V2.5-Pro-UltraSpeed, an advanced serving mode for its 1-trillion-parameter flagship AI model.
This new mode runs at over 1,000 tokens per second and peaked near 1,200 tokens per second during live demos. Xiaomi managed to pull this off using standard, commodity hardware—specifically a single 8-GPU node—without requiring any custom-built silicon. This hardware efficiency makes the MiMo model approximately 15 times faster than leading frontier systems like OpenAI’s ChatGPT and Anthropic’s Claude.
Inside the Microsecond Pipeline
The extreme speed increase relies on three synchronized technical layers working together. Xiaomi applied FP4 quantization selectively to the model’s Mixture-of-Experts layers. This reduces the intense memory bandwidth pressure that usually slows down large language models. They paired this with DFlash speculative decoding, a technique that predicts entire blocks of text in a single forward pass instead of generating one word at a time.
A custom TileRT runtime executes the entire pipeline at a microsecond scale. The resulting architecture achieves unprecedented speeds on commodity GPUs, officially rendering it 15 times faster than ChatGPT and Claude. Developers can test the raw speed immediately. Xiaomi opened a limited API trial for UltraSpeed running from June 9 to June 23, 2026. Access costs three times the standard per-token rate, and the company also uploaded the open-source weights to Hugging Face for local deployment.
Why Wait Times Define the Next AI Race
Raw inference speed changes how developers actually use artificial intelligence. Currently, complex multi-step agentic workflows—where an AI writes code, tests it, finds an error, and rewrites it—can take up to 20 minutes to complete using standard models like Claude Opus. UltraSpeed compresses that exact same autonomous feedback loop into just three to five minutes.
This reduction in idle time is pushing early testers to abandon traditional slow models for repetitive coding tasks. When an AI agent can iterate and correct its own mistakes 15 times faster than a human can read the output, the utility of the tool shifts from a conversational chatbot into a high-speed computational engine.