Xiaomi announced the MiMo-V2.5-Pro-UltraSpeed mode on June 8, 2026, developed in collaboration with TileRT, hitting a decoding speed of 1000 tokens per second on a 1-trillion-parameter model for the first time [1, 2]. The mode uses general-purpose GPUs instead of custom chips, achieving this speed efficiently with widely available hardware [2].
The new UltraSpeed mode offers about 10 times the generation speed of the previous MiMo-V2.5-Pro, though at three times the price [1]. Xiaomi designed it for enterprises and professional developers requiring faster AI inference on large models.
Starting June 9 through June 23, 2026, Xiaomi opened a limited API trial period for the MiMo-V2.5-Pro-UltraSpeed, prioritizing approved users in relevant fields [1]. The technical team highlighted the product launch on the evening of June 8, marking a milestone in decoding speeds without specialized chips [2].
The MiMo-V2.5-Pro-UltraSpeed mode targets faster and more efficient AI model deployment, providing significant improvements over Xiaomi’s prior offerings. The trial window will determine broader accessibility and user feedback.
The trial ends June 23, 2026, after which Xiaomi will review results before wider rollout.
