Iris-mini and Iris-pro are the strongest open-weight search agents in their class

Details

AllSpark, a Chinese AI lab, has released two open-source search agents — Iris-mini (35B parameters, built on Qwen3.6-35B-A3B) and Iris-pro (397B parameters, built on Qwen3.5-397B-A17B). Both work with a 256,000-token context window and achieve state-of-the-art results among open-weight models in their respective size classes on established search agent benchmarks. The training methodology is notably original. Instead of manually curating question-answer pairs, AllSpark reverse-engineers tasks from the web's link graph: starting from a seed page and its outgoing links, the system builds a graph of connected terms and generates multi-step questions requiring several reasoning hops. Intermediate terms are paraphrased so models cannot resolve queries via simple text matching — they must genuinely reason. Only tasks that a reference model fails without tools but solves with search access enter the dataset. Training alternates between supervised fine-tuning and reinforcement learning in a process the authors call SFT-RL climbing. A judge model — run on AllSpark's own large Qwen infrastructure rather than external APIs — filters training paths for correctness, efficiency, and absence of repetition loops. The hardest solved tasks feed back into each subsequent cycle. Beyond search benchmarks, the paper reports performance gains on general tool use and office work tasks, suggesting the reasoning discipline transfers across domains. The full training recipe is open-sourced alongside the models.