Iris-mini and Iris-pro: New Search Agents Top Deep Search Benchmarks

Iris is an intelligent search agent trained by researchers, available in two versions: Iris-mini with 35B active parameters and Iris-pro with 397B active parameters. It is designed to solve multi-step search tasks requiring evidence gathering and reasoning.

According to the paper, Iris outperforms all open-source search agents in its class, achieving 88.6% on BrowseComp and 92.9% on DeepSearchQA (for Iris-pro). Results are very close to the best closed systems.

It's a procedure alternating between supervised fine-tuning (SFT) and reinforcement learning (RL). The hardest solved and most efficient rollouts from each RL round are fed back into the next SFT pass, allowing the agent to progressively climb task difficulty.

Yes, researchers plan to release the weights and full training recipe, enabling teams to download and run the model locally or on their own infrastructure, which is important for organizations requiring data privacy.