Xiaomi pushes 1T-parameter model past 1,000 tokens/s

Xiaomi's MiMo team and inference-systems group TileRT announced that MiMo-V2.5-Pro-UltraSpeed has passed 1,000 tokens/s of output on a 1T-parameter flagship model. They describe this as the first time a model of that size has reached the mark. The figures come from Xiaomi and have not been independently verified.

The post is titled as a release of three "MiMo-V2.6 Series" models, but the text describes only the V2.5-Pro UltraSpeed mode. It gives no details on any V2.6 models.

Xiaomi says the speed comes from designing the model and serving system together, not from dedicated hardware. The company contrasts its approach with Cerebras' wafer-scale chips and Groq's SRAM-based chips. On the model side, the team quantized to FP4 to cut model size and memory traffic, which are the main bottleneck on general-purpose GPUs. It also used a speculative decoding method called DFlash. The source text cuts off before explaining the rest of the technical stack.

Access ran as an application-only API trial from June 9 to June 23, 2026. It cost three times the standard MiMo-V2.5-Pro rate for what Xiaomi said was about 10 times the output speed. Token Plan subscriptions were not supported. Approved users also got free chat access, capped at 10 queue entries a day and 30-minute sessions. Xiaomi says more than 3,000 companies and developers applied within 12 hours of the announcement on X, from industries including law, finance, telecoms, logistics and car manufacturing.

Xiaomi says the trial has ended and a commercial version is "coming soon." It has not announced pricing or a launch date.