MiMo V2.6 Pro UltraSpeed is a latency-focused variant within Xiaomi's MiMo-V2.6 series, launched alongside the V2.6-Pro flagship and the V2.6-Flash efficiency model. According to Xiaomi's update log and V2.6 release page, UltraSpeed preserves V2.6-Pro's flagship reasoning quality while delivering output up to roughly twenty times faster, positioning it for real-time interaction and other response-time-sensitive professional workflows rather than offline batch processing. The broader V2.6 series is built around natively omnimodal models trained with an "RSI" approach that scales reinforcement-learning compute on verifiable, complex tasks to expand the capability frontier through exploration and feedback.
UltraSpeed inherits the V2.6-Pro foundation, a sparse mixture-of-experts checkpoint with about one trillion total parameters and around forty-two billion active per token, combining a hybrid sliding-window and global attention layout with multi-token prediction to keep memory traffic low on long sequences. That underlying strength is what UltraSpeed compresses into a faster inference path, with Xiaomi describing the goal as preserving flagship Pro performance at substantially reduced latency for streaming and agentic use cases. Practically, it fits teams that want V2.6-Pro-class reasoning in interactive settings, such as live coding assistants, agent loops and other real-time applications, where response speed matters more than the lower per-token cost of the standard Pro or Flash endpoints.
