Xiaohongshu open-sources Iris: 35B search agent nears trillion-param models on BrowseComp

小红书技术REDtech · wechat · 2026-09-14

Xiaohongshu's AllSpark team open-sourced Iris, a search agent scoring 82.2 on BrowseComp at just 35B parameters (near trillion-param Kimi-K2.6) and 88.6 for the 397B Iris-pro, topping open-source on BrowseComp, DeepSearchQA and HLE. Key recipe: auto-generating training questions from hyperlink structures with entity paraphrasing to block string-matching shortcuts; SFT-RL alternating training with in-cluster reward judging/summarization and resumable long trajectories; plus surprising generalization to General ToolUse and Cowork tasks. Weights, eval code and paper released.

Original post →

More from coding & agent

coding & agent channel →