Xiaohongshu Releases New Large Model Architecture PIPO
Original title:小红书发布大模型新架构 PIPO ,让模型一次「吞两个、吐两个」
AI summary
Xiaohongshu proposed the PIPO architecture, which uses an input-side compressor to fold two tokens into one latent and an output-side MTP head to expand hidden states into additional tokens, halving input length and doubling output per step. Based on Qwen3.5-4B/9B backbones, it delivers up to +7.15 pass@4 improvement on benchmarks such as AIME 2025. In deployment tests, TTFT is accelerated by about 1.23× and TPOT by about 1.86×. Training uses two stages, SFT and On-Policy Distillation, distilling verifier verification capability into a lightweight confidence head.
Key Takeaways
- Xiaohongshu proposed the PIPO architecture,which uses an input-side compressor to fold two tokens…
- Based on Qwen3.
- 5-4B/9B backbones,it delivers up to +7.
- 15 pass@4 improvement on benchmarks such as AIME 2025.
- In deployment tests,TTFT is accelerated by about 1.
Quick Facts
- Published
- Updated
- Source
- 公众号:小红书技术(dots.llm)
- Category
- Papers & Research
- AI score
- 70
PIPO把输入压缩和输出多token预测统一起来,并且用蒸馏把验证成本前置,长推理的场景下这可能是下一代推理引擎的样子,做模型加速的应该认真看看。
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: Xiaohongshu Releases New Large Model Architecture PIPO. July 10, 2026. https://new.nat6.net/en/news/cmrerlfrs0046ihm87smc2el4/