Xiaomi's MiMo team and the inference systems team TileRT jointly announced that MiMo-V2.5-Pro's UltraSpeed mode has pushed a trillion-parameter (1T) flagship model past the 1000 tokens/s output mark for the first time. More importantly, this isn't a dumbed-down Flash variant — it's the full-strength Pro version.

That means roughly a 10x speedup while maintaining top-tier intelligence.

Real-World Speed

Take a complex visual dashboard generation task as an example:

• UltraSpeed version: done in 13 seconds
• Standard version: 6 minutes 15 seconds
• Up to a 28x speedup for the same quality of output

In hands-on testing, peak speed even reached 1426 tokens/s, outputting 25624 tokens in 32 seconds and generating 1000 lines of code. A Snake game takes 10 seconds to generate, and a macOS system UI can be reproduced in 1 minute.

How It Works

Unlike dedicated-hardware approaches such as Cerebras' wafer-scale integration or Groq's custom pure on-chip SRAM chips, Xiaomi pulled this off on generic GPUs — a single standard 8-GPU node.

The core technology has three parts:

1. FP4 quantization: a dramatic slim-down without losing accuracy

• Only the MoE experts get FP4 quantization; other modules keep their original precision
• Through FP4 quantization-aware training (QAT), the model's overall capability stays essentially on par with the original
• Model size and memory-access overhead are both dramatically reduced

2. DFlash speculative decoding: confirming multiple text segments in one go

• Uses a block-level masked parallel prediction method, removing the serial constraint of autoregressive drafting
• In coding scenarios the average acceptance length reaches 6.30, with 6-7 of every 8 drafted tokens accepted per verification round
• The draft model uses sliding-window attention (SWA), making per-prediction compute constant

3. TileRT custom compiled kernels

• Resident kernel engine: the compute pipeline lives inside the GPU and keeps flowing continuously
• Heterogeneous pipeline collaboration: communication, data movement, and tensor compute are finely decomposed at the tile level
• Microsecond-level hardware-software convergence: purpose-built for FP4 mixed quantization and DFlash

Pricing and API

The API is live with a limited-time trial price:

• Priced at 3x MiMo-V2.5-Pro, in exchange for roughly a 10x output speedup
• Based on MiMo-V2.5-Pro pricing, that works out to roughly ¥18 per million output tokens
• API trial only; Token Plan not supported yet
• Applications are open June 9 through June 23; approved users get two weeks of free limited-time Chat access

Why Speed Matters in Practice

The speedup isn't just "faster" — it unlocks entirely new usage patterns:

• Agent scenarios: if you estimate a task takes one minute, you'll watch it through to the end. If it takes five minutes, you'll probably go do something else and waste time coming back. A 10x speedup directly changes your working rhythm
• Sub-Agent concurrency: spin up one or two hundred sub-agents at once; with a 10x speedup and zero loss in model capability, the difference in experience is dramatic
• Real-time decision loops: trillion-parameter models can plug into time-critical scenarios like high-frequency quantitative trading signal generation and instant fraud-detection interception

How to Apply

• API application: https://platform.xiaomimimo.com/ultraspeed
• Chat trial: https://ultraspeed.xiaomimimo.com
• Open-source weights (FP4-DFlash): https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash

Tip: The current high acceptance rates are still concentrated mainly in structured tasks like coding; general conversation still has room for optimization. Inference capacity is tight, so broad commercial availability will take time.