AllSpark releases open-weight search agents Iris-mini and Iris-pro built on Qwen

Chinese lab AllSpark has released two open-weight agents that research the web on their own, together with a full training recipe. Built on Qwen models, Iris-mini has 35 billion parameters and Iris-pro 397 billion; both use a 256,000-token context window. The team says they lead open-weight search agents in their size classes. With context management on, Iris-mini scored 82.2 on BrowseComp and Iris-pro 88.6.

Training questions are built backward from the link structure of web pages, and a stronger teacher model's solution paths go through two filtering rounds. The team argues that how context is managed in long sessions can matter more than the reported gaps between models. Context management lifted the smaller model's BrowseComp score by up to 21.2 points. The paper also reports gains on untrained tasks such as general tool use and office work.