Iris-pro scores 92.9 on DeepSearchQA with a single ReAct agent

Iris-pro, a 397B parameter search agent developed by an open-source research team, achieved a 92.9 accuracy score on the DeepSearchQA benchmark with context management enabled, setting a new state-of-the-art result for open-source search agents. The agent, along with Iris-mini (35B-A3B), was trained at massive scales and demonstrates improved multi-hop reasoning and retrieval performance, enabling production teams to deploy more reliable search agents without proprietary systems.

arXiv https://arxiv.org/abs/2609.04304 Iris-pro scores 92.9 on DeepSearchQA with a single ReAct agent Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated. Iris-pro, a 397B parameter search agent, achieves 92.9 accuracy on DeepSearchQA with context management enabled, setting a new benchmark for open-source search agents. This pushes the frontier of multi-hop reasoning and retrieval performance, enabling production teams to deploy more reliable and efficient agents for complex search tasks without requiring proprietary systems or costly test-time verification. Two search agents, Iris-mini and Iris-pro, were trained at massive scales 35B-A3B and 397B-A17B and achieved state-of-the-art results among open-source search agents, reaching scores of up to 92.9 on certain benchmarks, which enables shipping more accurate and capable search agents in production environments.