Chinese research lab AllSpark has released Iris-mini and Iris-pro, two open-weight search agents based on the Qwen model family. According to a research paper published by the team, Iris-mini features 35 billion parameters while Iris-pro contains 397 billion parameters, both operating with a 256,000-token context window. The models achieved leading benchmarks among open-weight search agents in their respective size classes, while also demonstrating unexpected performance gains on general tool use and office tasks.
To train the agents, AllSpark developed a dataset generation process built backward from web page link structures. The method creates multi-step questions requiring chained reasoning steps, replacing key terms with paraphrases so queries cannot be solved via simple text search. Candidate solution paths generated by a teacher model underwent two filtering rounds before the models were optimized through reinforcement learning against a live web search. The lab utilized an alternating framework termed “SFT-RL climbing” to continuously feed efficient solution paths back into training cycles.
The authors also addressed benchmark evaluation methodologies, demonstrating that runtime context management strategies often skew performance comparisons. To isolate model capability from external scaffolding, AllSpark evaluated benchmarks with and without context management while standardizing tools and context limits, testing systems as single agents without secondary verification steps.
Why it matters
Open-weight search agents now offer high-parameter reasoning capabilities, reducing dependency on proprietary web search APIs.
The open SFT-RL training pipeline provides a replicable recipe for enterprise teams building custom autonomous research agents.
Source: the-decoder.com
