Iris-mini and Iris-pro are the strongest open-weight search agents in their class
decoder ·
The AllSpark team has released Iris-mini and Iris-pro, two open-source search agents, along with a full training recipe. According to the paper, Iris-mini has 35 billion parameters and Iris-pro has 397 billion; both are built on Qwen-series models, use a 256,000-token context window, and are claimed to be the strongest open-weight search agents in their size classes. The training pipeline reverse-engineers multi-step questions from web link graphs, replaces clues with paraphrases, and filters generated solution paths through correctness checks and a judge model before reinforcement learning against live web search. The team says runtime context management can have a larger impact on benchmark results than differences between models, so they test with and without it while holding tools, context limits, and judge model constant. On BrowseComp, BrowseComp-ZH, DeepSearchQA, and Humanity's Last Exam, Iris-mini scores 82.2, 84.8, 86.9, and 52.3, while Iris-pro scores 88.6, 85.1, 92.9, and 56.4 with context management enabled. Iris-mini leads its class on three of four benchmarks and beats XYZ-Aquila-mini on BrowseComp by 3.4 points, while Iris-pro leads or ties in the larger class. Context management helps the smaller model more, boosting BrowseComp by up to 21.2 points, and the paper notes that Iris-mini needs more steps and hits the context limit more often. The authors also report that the training data and models improved performance on unseen tasks such as general tool use and office work, and they provide model weights on Hugging Face plus code on GitHub.
decoder