Xiaomi introduced UltraSpeed mode for its MiMo-V2.5-Pro language model, a software update that breaks the 1,000 tokens-per-second generation barrier on standard GPUs. The company positions this release as an "ultimate co-design" of the underlying system and the 1-trillion-parameter model. Developers can now access significantly faster output speeds through a new API tier.
New API tier delivers ten times faster output for enterprise developers at triple standard pricing
The MiMo-V2.5-Pro architecture serves as the foundation for this performance jump. Xiaomi states that UltraSpeed mode delivers roughly ten times faster output compared to standard API access on the same hardware. The update targets enterprise users and professional developers who require rapid text generation without upgrading their physical GPU infrastructure.
UltraSpeed pricing operates at three times the standard rate for all token categories. Input tokens cost 3 yuan per million during a cache miss, while output tokens run 6 yuan per million. Cache hit inputs drop to 0.025 yuan per million under normal conditions, but UltraSpeed applies the triple multiplier across the board.
The vendor restricts this new tier to API access only and does not support it through Token Plans. Xiaomi notes that supply remains constrained during the trial phase, which runs from June 9 to June 23, 2026. The company explicitly states there is no guaranteed approval timeline or success rate for applications.
Trial access prioritizes enterprise accounts and professional developers over general consumers. Developers must submit formal applications to secure a spot within the two-week window. The trial period concludes on June 23, 2026, after which standard pricing tiers resume without UltraSpeed availability.
Xiaomi confirms that the new mode achieves its speed claims through system-level optimization rather than raw hardware upgrades. The trial phase allows selected users to test the performance gains before any potential broader rollout.
Source: GIZMOCHINA
