Iris-mini and Iris-pro are the strongest open-weight search agents in their class
AllSpark released Iris-mini (35B) and Iris-pro (397B) as open-source search agents built on Qwen-series models, with full training recipes and code on Hugging Face and GitHub. Training data is reverse-engineered from web link structures, generating multi-step reasoning tasks where paraphrasing prevents simple text-search solutions, ensuring genuine reasoning over lookup. A two-stage filtering pipeline using a teacher model and a data-derived judge model, combined with SFT-RL climbing against liv
Analysis
TL;DR
- AllSpark released Iris-mini (35B) and Iris-pro (397B) as open-source search agents built on Qwen-series models, with full training recipes and code on Hugging Face and GitHub.
- Training data is reverse-engineered from web link structures, generating multi-step reasoning tasks where paraphrasing prevents simple text-search solutions, ensuring genuine reasoning over lookup.
- A two-stage filtering pipeline using a teacher model and a data-derived judge model, combined with SFT-RL climbing against live web search, produces the final agents.
- Context management during inference has a dramatically larger impact on benchmark scores—especially for smaller models—raising questions about how much of reported performance reflects model capability versus scaffolding.
- Search training yields unexpected generalization benefits, improving performance on unrelated tool use and office work tasks, suggesting search may be a foundational rather than narrow skill.
Why It Matters
This release provides one of the most complete open-source packages for building search agents to date, including models, training data construction methodology, evaluation harness, and reinforcement learning pipelines. It also surfaces a critical methodological concern: many published search agent benchmarks may conflate model quality with context-management tricks, which has direct implications for how the community evaluates and compares agentic systems. The unexpected cross-domain improvements from search training further challenge the assumption that search is a narrow capability.
Technical Details
- Model architecture: Iris-mini (35B parameters, based on Qwen3.6-35B-A3B) and Iris-pro (397B parameters, based on Qwen3.5-397B-A17B), both supporting a 256,000-token context window.
- Training data construction: Tasks are generated by traversing web page link graphs from seed pages, creating multi-step questions that require chaining reasoning across connected terms. Every intermediate term is paraphrased to prevent trivial text search, and a reference model filter ensures only questions solvable with tools but not without them are included.
- Two-stage filtering and SFT-RL climbing: A stronger teacher model generates solution paths (reasoning + queries + results), which are filtered for correctness, repetition loops, and search depth. A judge model with data-derived criteria performs step-by-step review. Supervised fine-tuning and reinforcement learning alternate iteratively, with the hardest solved tasks and most efficient paths feeding back into each cycle. All inference runs internally on the team's own Qwen model, avoiding external service dependency.
- Context management strategy: The Iris Harness includes history-discard techniques and a second-attempt mechanism where failed runs are condensed into notes recording what was checked and ruled out, appended to subsequent attempts. This significantly boosts smaller models by mitigating faster context consumption.
- Benchmark results (with context management enabled): Iris-mini scored 82.2 on BrowseComp, 84.8 on BrowseComp-ZH, 86.9 on DeepSearchQA, and 52.3 on Humanity's Last Exam. Iris-pro scored 88.6, 85.1, 92.9, and 56.4 respectively. Context management boosted Iris-mini's BrowseComp score by up to 21.2 points.
Industry Insight
- Benchmark reporting standards for search agents need revision: the dramatic gap between scored results with and without context management suggests the community should mandate paired evaluations that isolate model capability from engineering scaffolding, or risk inflating perceived progress.
- The cross-task generalization from search training (tool use, office work) suggests that investing in search as a foundational agent skill could yield broader ROI than treating it as a standalone module, influencing how teams allocate training budgets and design curricula.
- The open release of training recipes, data construction pipelines (planned), and an evaluation harness compatible with any OpenAI-endpoint lowers the barrier for independent replication and comparison, likely accelerating the open-weight search agent ecosystem and putting pressure on proprietary systems to match transparency.
Disclaimer: The above content is generated by AI and is for reference only.