Summary and commentary: Geek Haus. Editorial responsibility: Jay Jang, CEO

Xiaomi says MiMo-V2.5-Pro-UltraSpeed reaches 1,000+ tokens per second for a 1T-parameter AI model

Jun 9, 2026, 00:27·mimo.xiaomi.com

Geek Haus editorial team · Sources and editorial policy

EDITOR BRIEF

Xiaomi released MiMo-V2.5-Pro-UltraSpeed with TileRT, claiming the first 1T-parameter model to exceed 1,000 tokens per second in decode speed, reaching up to about 1,200 tokens per second. The API is offered in a limited trial from June 9 to June 23, 2026, at 3x the MiMo-V2.5-Pro price while promising roughly 10x faster generation.

COMMENTARY

The launch highlights how inference speed is becoming a key differentiator as large models move from batch-style outputs toward real-time collaboration and agentic workflows. Limited access and enterprise prioritization suggest high-speed inference remains resource-constrained, making infrastructure efficiency a competitive battleground.

COMMENTS

Discussion

> geekhaus:~$ read next

Cloudflare Blog

AI code reviewer cuts inference costs 68%, from $2.43 to $0.77 per review, after prompt-cache, call-count, and model changes

Hugging Face Blog

Reflection releases Beam, a 501B-parameter Apache 2.0 MoE model for coding agents with 1M-token context

research.meta.ai