AI BriefingMimo Tilert 1000tps
Key point
Xiaomi collaborated with TileRT to achieve a generation speed of over 1000 tokens per second on a model with 1 trillion parameters.
Details
Xiaomi has released MiMo-V2.5-Pro-UltraSpeed through collaboration with TileRT. Despite being a 1 trillion (1T) parameter scale model, it broke through industry limits by recording a decode speed of over 1000 tokens/s.
This technological innovation is expected to transform the AI usage paradigm beyond simply improving speed.
- Evolution of Reasoning Capability: Based on the fast speed, it can run parallel reasoning paths (Best-of-N, Tree Search) to self-correct errors in real time and add depth to its thinking.
- Maximized Productivity: It dramatically reduces the latency of Coding Agents, drastically increasing development efficiency.
- Real-time Decision Making: It can be applied to real-time scenarios requiring millisecond-level responses, such as high-frequency trading, fraud prevention, and medical surgery assistance.
The MiMo-V2.5-Pro-UltraSpeed API will only be available for a limited, application-based period (June 9 - June 23, 2026) due to limited high-speed inference resources. The price is 3 times higher than the existing MiMo-V2.5-Pro, but the generation speed is about 10 times faster.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Inception Releases Ultra-Fast LLM 'Mercury 2.5' Generating 770 Tokens Per Second