AllSpark Research has introduced Iris, a series of open-weight search agents comprising Iris-mini (35B total, 3B active) and Iris-pro (397B total, 17B active), both derived from the Qwen3.5/3.6 MoE series. To address the challenges of long-horizon, multi-hop web searching and context saturation, the researchers developed an iterative "SFT-RL climbing" training recipe. This method alternates between supervised fine-tuning (SFT) on teacher-generated trajectories—filtered for quality and turn-level noise—and reinforcement learning (RL) using group-relative policy gradients. Successful RL rollouts are fed back into the SFT dataset, automatically increasing task difficulty. Training data is generated via an automated pipeline that reverse-constructs multi-hop questions from web graph subgraphs, utilizing an "anchor abstraction" operator to prevent simple string matching. Notably, the study emphasizes that inference-time context management (CM) is critical; implementing a "discard-all" strategy, which clears tool history when context limits are reached, significantly boosts performance. On the BrowseComp benchmark, Iris-pro achieves a score of 88.6 with discard-all enabled, reaching 90.3 when combined with a "retry" strategy. These models achieve state-of-the-art results among open-source agents on benchmarks including BrowseComp, BrowseComp-ZH, DeepSearchQA, and Humanity's Last Exam (HLE), demonstrating that sophisticated training recipes and context management can allow open-weight models to rival proprietary systems in complex reasoning tasks.