Back to News
Xiaomi•June 8, 2026
MiMo-V2.5-Pro-UltraSpeed
Paid
AI language model
Xiaomi
TileRT
MiMo-V2.5-Pro-UltraSpeed
### TL;DR
MiMo-V2.5-Pro-UltraSpeed is an advanced AI language model developed by Xiaomi, capable of generating up to 1,200 tokens per second on a 1-trillion-parameter model. This performance is achieved through a deep collaboration with TileRT, focusing on extreme model-system co-design to optimize inference speed on commodity GPUs.
Key Insights & Metrics
Pricing
$1.305 per Million tokens Input (Cache Miss) & $ 2.61 per Million tokens (Output )
Cost structure
Version
V2.5-Pro-UltraSpeed
Current release version
Hardware
Requires an 8-GPU enterprise server node and features day-0 native optimization for AMD Instinct GPUs or equivalent Nvidia Tensor Core clusters. (Model can be accessed by API as well)
Compute requirements
Category
Paid
Licensing model
Region
China
Primary region
Key Features
- Generates up to 1,200 tokens per second on a 1-trillion-parameter model
- Achieves high inference speed on commodity GPUs through model-system co-design
- Utilizes FP4 quantization and DFlash speculative decoding for enhanced performance
- Available through a limited-time application-based API from June 9 to June 23, 2026
- Offers free Chat access to approved users during the trial period
- Enables real-time decision-making in critical scenarios
Discussion
0
Upvotes
0
Downvotes
0 reviews
Sign in to leave a review
Reviews
No reviews yet. Be the first to review!