Overview
Iris is a pair of search-agent models open-sourced by Xiaohongshu's AllSpark team on September 14, 2026: the 35B Iris-mini and the 397B Iris-pro, built on Qwen3.6-35B-A3B and Qwen3.5-397B-A17B respectively, both with 256K context. The team focused on one job. Training data is generated backwards from the hyperlink structure of the web: seed pages spawn an entity graph and multi-hop questions, then every entity except the answer is rewritten as a descriptive reference to erase the clues a model could simply copy. A sample survives only if the model cannot answer with the web closed and can answer correctly and uniquely with it open.
Evaluation covers BrowseComp, BrowseComp-ZH, DeepSearchQA and HLE. Iris-pro scores 88.6, 85.1, 92.9 and 56.4; Iris-mini scores 82.2, 84.8, 86.9 and 52.3, with its 82.2 on BrowseComp approaching models many times its size. Both sit at the front of open models in their parameter class.
Key Features
- Two sizes for different deployment budgets: Iris-mini at 35B (Qwen3.6-35B-A3B base) and Iris-pro at 397B (Qwen3.5-397B-A17B base), covering local deployment and maximum capability respectively
- Backward-generated training questions: Questions are built from web hyperlink structure into entity graphs and multi-hop queries, with every non-answer entity rewritten as a descriptive reference so the prompt leaks fewer clues
- Closed-book and open-book filtering: Only samples the model fails closed-book but answers correctly and uniquely open-book stay in the training set, ensuring retrieval is genuinely required
- SFT-to-RL climbing recipe: Training climbs from supervised fine-tuning to reinforcement learning, with scoring and summarization kept inside the training cluster instead of calling external APIs, and ultra-long trajectories that can resume after a break
- 256K context: Both versions support 256K context, suited to multi-round retrieval and synthesis across long source material
- Gains spill over to general tool use: The team observed search training paying off beyond search, with improvements on BFCL, τ-bench, OfficeQA and APEX
Use Cases
- Deep research Q&A that needs multi-hop retrieval and source checking
- Adding reliable web search to an existing agent, replacing a general model's default browsing behavior
- Chinese-language retrieval, one of the benchmarks Iris targets most directly
- Deploying a search agent locally or in a private environment, where the 35B version is easier on VRAM
- Pairing with a general LLM, letting Iris handle retrieval and source gathering while the larger model writes and decides
Pros
- Both versions come from the same training recipe, and the small one approaches trillion-parameter models on BrowseComp
- 256K context covers long-source synthesis without constantly truncating retrieved material
- Scoring and summarization run inside the training cluster and long trajectories resume after breaks, which is easier to operate
- Weights, paper and model cards are fully open, ready to download or fine-tune
- Search gains spill over into general tool use rather than showing up on a single benchmark
Pricing
Weights are released as open source and can be self-hosted; inference cost depends on the version and the hosting platform. The two sizes map to different VRAM and throughput budgets.
Summary
Iris suits teams that treat search as a core capability: deep research Q&A, agents that need dependable retrieval, or Chinese-language scenarios where answers must be verifiable. The 35B version is friendlier to local deployment while the 397B version targets the ceiling. If you only need occasional web answers, a general model's built-in browsing is usually enough and a dedicated search model is not worth the extra moving part.
Version History
- Xiaohongshu's AllSpark open-sources the Iris search-agent models in 35B and 397B sizes (2026-09-14): Xiaohongshu's AllSpark team open-sourced the Iris search-agent family: Iris-mini at 35B (Qwen3.6-35B-A3B base) and Iris-pro at 397B (Qwen3.5-397B-A17B base), both with 256K context. On four benchmarks, Iris-pro scores 88.6 on BrowseComp, 85.1 on BrowseComp-ZH, 92.9 on DeepSearchQA and 56.4 on HLE; Iris-mini scores 82.2, 84.8, 86.9 and 52.3, with its 82.2 on BrowseComp approaching trillion-parameter models. Training data is generated backwards from web hyperlink structure, rewriting every non-answer entity as a descriptive reference, and filtered so samples are unanswerable closed-book yet uniquely answerable open-book. Training uses an SFT-to-RL climbing recipe with scoring and summarization inside the training cluster and resumable ultra-long trajectories. Search gains also spill over to general tool use and office benchmarks including BFCL, τ-bench, OfficeQA and APEX. The paper is arXiv:2609.04304, with weights and model cards on GitHub and Hugging Face.