Xiaomi MiMo and TileRT jointly release UltraSpeed mode, 1T model output exceeds 1000 tokens/s
Original title:Xiaomi MiMo 携手 TileRT|1T 模型首次突破 1000 tokens/s 输出速度
AI summary
Xiaomi MiMo and TileRT jointly released the MiMo-V2.5-Pro-UltraSpeed mode, making the 1T-parameter flagship model's output speed exceed 1000 tokens/s for the first time. On the model side, it uses FP4 mixed quantization (quantizing only MoE Experts) and DFlash block-level masked parallel speculative decoding (average accepted length 6.30 tokens in coding scenarios); on the system side, TileRT introduces a persistent kernel engine and heterogeneous pipeline collaboration. The API is open for a limited time (June 9–23, 2026), priced at 3x MiMo-V2.5-Pro, with about 10x speed improvement. FP4 weights and DFlash model checkpoints have been open-sourced on HuggingFace.
Key Takeaways
- Xiaomi MiMo and TileRT jointly released the MiMo-V2.
- 5-Pro-UltraSpeed mode,making the 1T-parameter flagship model's output speed exceed 1000 tokens/s…
- On the model side,it uses FP4 mixed quantization (quantizing only MoE Experts) and DFlash…
- 30 tokens in coding scenarios)
- on the system side,TileRT introduces a persistent kernel engine and heterogeneous pipeline…
Quick Facts
- Published
- Updated
- Source
- 公众号:小米 MiMo
- Category
- Model Releases
- AI score
- 80
万亿模型首次在通用GPU上突破1000 tokens/s,不是专用硬件的胜利而是模型与系统Codesign的胜利,做实时AI应用的都应该盯紧这一套方案。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: Xiaomi MiMo and TileRT jointly release UltraSpeed mode, 1T model output exceeds 1000 tokens/s. June 9, 2026. https://new.nat6.net/en/news/cmq63dzew04srsl5ix1r8n4ne/