AllSpark releases Iris-mini and Iris-pro open-weight search agents with a full training recipe
The 35B and 397B Qwen-based models build training questions backward from the web's link structure and reportedly improve on tool use and office tasks they never trained for.
The AllSpark team, a Chinese lab, has released two open-source search agents, Iris-mini and Iris-pro, along with a full training recipe. Iris-mini has 35 billion parameters and builds on Qwen3.6-35B-A3B; Iris-pro has 397 billion parameters and builds on Qwen3.5-397B-A17B. Both work with a 256,000-token context window. According to the team's paper, each delivers the strongest results among open-weight search agents in its size class. Search agents built on language models research the web on their own: they must understand a question, decide what to search for, interpret results, and judge when they have enough evidence. How much of that leading systems actually do is debated, since on established benchmarks they mostly use the web to confirm knowledge already absorbed during training. AllSpark's training pipeline constructs tasks backward from the link structure of the web, reverse-engineering questions whose answers require following chains of pages. The paper reports that the training data and models also improved performance on tasks they were never trained for, including general tool use and office work. The release matters for developers who want a capable retrieval agent without depending on closed APIs, and the published recipe lets other groups reproduce or adapt the approach.