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

The AllSpark team has introduced Iris-mini and Iris-pro, a pair of open-source search agents constructed on top of Qwen language models. The release targets developers looking for capable, open-weight alternatives to proprietary search and retrieval tools.

According to the team's paper, both models top benchmark rankings among open-weight competitors within their respective size categories. The researchers also noted that the training methodology and datasets yielded performance improvements on tasks the models were not explicitly trained to handle.

This matters because open-weight agents give the community greater transparency and flexibility for building custom retrieval pipelines. Strong benchmark results from a smaller open model could lower the barrier to entry for organizations that lack the compute budget for large proprietary systems.

The generalization to untrained tasks suggests the training data may have improved underlying reasoning or search behaviors rather than simply overfitting to benchmark formats. If independent evaluations confirm these claims, Iris could become a viable baseline for future open-source agent development.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.