Chinese lab AllSpark has dropped two new open-weight search agents, Iris-mini and Iris-pro, claiming top-tier performance in their respective size classes. The 35-billion-parameter Iris-mini and the massive 397-billion-parameter Iris-pro are built on Qwen-series foundations. But the real story isn’t just the parameter count. It’s how they were trained. The team reverse-engineered questions from web link structures, forcing the models to chain reasoning steps rather than rely on simple text matches. If a reference model could solve the task without tools, it was tossed out. This ensures the dataset only contains problems that actually require search.
The training pipeline is rigorous. A teacher model generates solution paths, which then face a two-stage filtering process. First, a check for correctness and loops. Second, a step-by-step review by a judge model whose criteria were derived from the data itself, not hand-coded rules. The authors call this iterative improvement “SFT-RL climbing.” Crucially, the judge model and result summaries run inside the training cluster on AllSpark’s own large Qwen model. This keeps the process self-contained and independent of external APIs. It’s a closed loop designed to refine efficiency and accuracy simultaneously.
Benchmark results show Iris-mini scoring 82.2 on BrowseComp, beating the next-best model, XYZ-Aquila-mini, by 3.4 points. Iris-pro hits 88.6 on the same test. However, the paper highlights a critical variable: context management. When the team isolated this factor, they found it boosted Iris-mini’s BrowseComp score by up to 21.2 points. The smaller model consumes context faster because it takes more steps to reach an answer. This suggests that runtime scaffolding often masks true model capability. Many reported gains in the industry may come from clever history discarding rather than raw intelligence.
On Humanity’s Last Exam, a test of expert-level academic knowledge, gains were modest. Web search plays a supporting role there, so the agents’ search prowess matters less. The best results came from combining history discarding with a second attempt at the problem. This raises a skeptical question for the industry: are we measuring better models or better memory management? AllSpark’s transparency about this distinction is refreshing. It forces a reevaluation of how we benchmark autonomous agents. If context tricks drive scores, the race for larger models might be missing the point entirely.
