Iris-mini and Iris-pro set the standards for open-weight search agents.

AllSpark released its open-weight web search agents built upon Qwen, Iris-mini and Iris-pro. They are capable of autonomously browsing the web and gathering information. They are trained on Qwen3.6-35B-A3B (Iris-mini; 35 billion params) and Qwen3.5-397B-A17B (Iris-pro; 397 billion params), with context length 256,000 tokens.

Iris-mini achieved a score of 82.2 on BrowseComp, a benchmark for hard web search. Iris-pro achieved a score of 88.6. They also achieved scores of 84.8 and 85.1 on BrowseComp-ZH (Chinese), 86.9 and 92.9 on DeepSearchQA, and 52.3 and 56.4 on Humanity’s Last Exam (expert knowledge benchmark).

Search related tasks generalize well to other tasks not seen during training such as tool use and office tasks. Iris-mini outperforms SOTA on BrowseComp by 3.4 points (Iris-mini vs XYZ-Aquila-mini).

You can get the model weights from Hugging Face, the code from GitHub, and it’s open source.

Note that Iris-pro (397B) is better deployed in datacenters instead of your personal workstation. Iris-mini is somewhat reasonable, but you still need better computing resources than laptops. Now with a GPU cluster, you have your own open-source web search engine powered by local data without relying on proprietary APIs!