Xiaomi · 2026-06-08 · major
Xiaomi Ships MiMo-V2.5-Pro-UltraSpeed — FP4-Quantized 1.02T MoE Hits ~1,200 Tokens/Sec on Stock 8-GPU Nodes via Block-Diffusion 'DFlash' Speculative Decoding, MIT-Licensed FP4-DFlash Checkpoint Lands on Hugging Face
Xiaomi cranks MiMo-V2.5-Pro to four-figure generation speed with MXFP4 expert quantization and a block-diffusion speculative decoder, releasing the full FP4-DFlash checkpoint under MIT and opening a free public trial June 9–23.
An FP4 backbone plus a block-diffusion drafter pushes Xiaomi's 1T MiMo MoE past 1,000 tokens per second on a stock 8-GPU node.
Key specs
| Parameters | 1.02T |
|---|---|
| Active params | 42B |
| Context window | 1M tokens |
| Drafter params | 5-layer BF16 |
| Throughput tokens per sec | 1000–1200 |
| Hardware | 8 GPUs |
| Trial price multiplier | 3× MiMo-V2.5-Pro API |
What is it?
MiMo-V2.5-Pro-UltraSpeed is an inference-optimised variant of Xiaomi's flagship MiMo-V2.5-Pro mixture-of-experts model — 1.02T total / 42B active parameters with a 1M-token context. The release pairs the existing MoE backbone with a 5-layer BF16 drafter for speculative decoding, ships the resulting FP4-DFlash checkpoint on Hugging Face under MIT, and opens a paid API trial during June 9–23 at 3× the cost of standard MiMo-V2.5-Pro.
How does it work?
Two pieces stack. First, MXFP4 (Microscaling FP4) quantization is applied only to the MoE expert weights while attention and the drafter stay at higher precision, cutting expert memory roughly 4× without measurable quality loss. Second, 'DFlash' is a block-diffusion speculative decoder: the drafter proposes an 8-token block in one forward pass via masked parallel prediction, and the backbone verifies in a single step. Combined with Xiaomi's existing 6:1 sliding-window-to-global-attention layout, the system sustains 1,000–1,200 tokens/sec on a commodity 8-GPU node.
Why does it matter?
Trillion-parameter MoEs usually only hit triple-digit throughput on multi-rack hardware. Pushing past 1,000 tokens/sec on a single 8-GPU node — and releasing the recipe and weights — drops the cost of serving long-context agentic workloads and turns FP4 plus block-diffusion drafting into a reproducible open-source baseline that competing labs can attack.
Who is it for?
inference engineers, MoE researchers, self-hosters, agentic-systems teams
Try it
huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash