Iris-mini – open search agent on Qwen3.6-35B-A3B (github.com)

Iris-mini and Iris-pro have been introduced as open-weight search agents derived from the Qwen3.5/3.6 series, showcasing advanced capabilities in handling search processes. Specifically, Iris-mini, with 35 billion parameters, and Iris-pro, with 397 billion parameters, are designed to adeptly manage the complexities of search by intelligently deciding what to search, interpreting results, and determining when sufficient evidence has been gathered. These agents utilize a model architecture that includes innovative context management techniques, allowing them to retain coherent conversations over extended interactions. This announcement is significant for the AI/ML community as it presents notable advancements in open-source search technology. Both models demonstrate strong performance across various benchmarks, such as BrowseComp and DeepSearchQA, with Iris-mini achieving an accuracy of 82.2% in its main settings. The implications of this development extend to enhancing interactive AI systems, making them more efficient and powerful in real-world applications. The technology underlying Iris includes strategies like "discard-all" for context management, ensuring that the agent can begin fresh when faced with hard questions, thus improving overall search effectiveness. The collaborative effort and forthcoming data construction pipelines from the project reflect a commitment to open-source innovation in AI.