AllSpark releases Iris-mini and Iris-pro, two open-weight search agents built on Qwen
Researchers at the Chinese lab AllSpark have released Iris-mini and Iris-pro, two open-weight search agents that post the strongest scores in their respective size classes on four established web-research benchmarks. The team also published the full training recipe, the agent infrastructure, and all four evaluation benchmarks, with model weights on Hugging Face and code on GitHub.
What the Iris models are
Iris-mini has 35 billion parameters and Iris-pro has 397 billion. Both are built on Qwen-series base models: Qwen3.6-35B-A3B for the smaller model and Qwen3.5-397B-A17B for the larger one. Both work with a 256,000-token context window. According to the team, they deliver the strongest results among open-weight search agents in their size class on the benchmarks they tested.
Search agents are language models that research the web on their own. They read a question, decide what to search for, interpret the results, and decide when they have enough evidence to answer. How much of that work the model actually does, as opposed to scaffolding around it, is a recurring debate in the field.
How the training data is built
The training pipeline reverse-engineers research questions from the link structure of the web. Starting from a seed page and its outgoing links, the system constructs a graph of terms and relationships, then generates a multi-step question whose answer requires chaining several connected steps together. Every term except the final answer is replaced with a paraphrase, so no clue can be resolved by a simple text search. The agent has to reason rather than look things up.
Only questions that a reference model cannot solve without tools but can solve with the right sources make it into the dataset, which keeps the tasks both hard and clearly verifiable.
Two-stage filtering and SFT-RL climbing
A stronger teacher model generates solution paths made up of reasoning, search queries, and results. Those paths go through two rounds of filtering. The first checks the full path for correctness, repetition loops, and search depth. The second is a step-by-step review by a judge model whose criteria were derived from the data itself rather than set by hand.
After that, the model is improved through reinforcement learning against a live web search. The judge model and result summaries run inside the training cluster on the team’s own large Qwen model, so training does not depend on external services. Supervised fine-tuning and reinforcement learning alternate in a process the authors call SFT-RL climbing. The hardest solved tasks and the most efficient solution paths from each round feed back into the next training cycle.
Context management matters more than the model gap
The team argues that runtime context management on common benchmarks often makes a bigger difference than the reported gaps between systems. During long research sessions the context can fill up before the agent has resolved every sub-question. Tricks like discarding the conversation history extend the research artificially and say little about the model’s actual quality.
To isolate the effect, the team tests every benchmark with and without context management while keeping tools, context limits, and the judge model constant. Results reported only with management turned on cannot be cleanly split between what comes from the model and what comes from the scaffolding around it. The Iris scores come from a single agent with no helper agents and no extra verification steps at the end.
Benchmark scores
Testing covered BrowseComp, which tests the ability to find rare facts from indirect clues, its Chinese counterpart BrowseComp-ZH, DeepSearchQA, which evaluates the completeness of retrieved evidence, and Humanity’s Last Exam, which poses academic questions at expert level. With context management on, Iris-mini scores 82.2, 84.8, 86.9, and 52.3 across the four benchmarks. Iris-pro reaches 88.6, 85.1, 92.9, and 56.4.
Iris-mini leads its size class on three of four benchmarks and beats the next-best model, XYZ-Aquila-mini, on BrowseComp by 3.4 points, though it trails on DeepSearchQA. Iris-pro leads or ties in its larger class and sometimes approaches systems that need far more compute.
Why the smaller model gains more from context management
Context management has a much larger effect on the smaller model, boosting BrowseComp scores by up to 21.2 points. According to the paper, the cause is not a smaller token budget but faster consumption. Iris-mini needs more steps for the same tasks and hits the context limit more often. On Humanity’s Last Exam the gains are smaller because that benchmark leans more on domain knowledge and academic reasoning, where web search plays a supporting role. The best scores come from combining history discarding with a second attempt. If the first try fails, the system condenses it into a short note that records what was already checked and ruled out, and that note is appended to the task for the next run.
A case where the ground truth was wrong
In the paper’s appendix the team describes a case where the agent was marked wrong even though the answer matched the source material. A question in BrowseComp-ZH targeted the series Game of Thrones. The agent answered Bolton, but the ground truth said Lannister. The character in question, Sansa Stark, actually marries Ramsay Bolton in her second marriage. The agent’s answer was correct. The team writes that contradictions like this between ground truth and source material motivate better benchmarks.
Transfer to tasks the models were never trained for
Beyond search, the authors report an unexpected side effect. Both the generated training data and the specialized models improved performance on tasks they were never trained for, including general tool use and office work. The team suggests that search may function more as a foundational skill than a narrow specialty, since the learned behavior helps wherever an agent has to work with incomplete information.
What is released now and what comes next
The model weights for Iris-mini and Iris-pro are available on Hugging Face, and the code is on GitHub. The current release includes Iris Harness, which contains the agent loop, tools, context management strategies, and all four benchmarks with evaluation. The team plans to release the data construction and training pipelines later.
FAQ
What are Iris-mini and Iris-pro?
Iris-mini and Iris-pro are open-weight search agents released by AllSpark. Iris-mini has 35 billion parameters and is built on Qwen3.6-35B-A3B. Iris-pro has 397 billion parameters and is built on Qwen3.5-397B-A17B. Both use a 256,000-token context window and post the strongest open-weight scores in their size class on BrowseComp, BrowseComp-ZH, DeepSearchQA, and Humanity’s Last Exam.
How is the training data generated?
The pipeline reverse-engineers research questions from the link structure of web pages. Starting from a seed page and its outgoing links, the system builds a graph of terms and relationships, then generates a multi-step question whose answer requires chaining several connected steps. Every term except the final answer is paraphrased, so no clue can be resolved by a simple text search.
Where can the models and code be downloaded?
The model weights for Iris-mini and Iris-pro are available in a collection on Hugging Face. The Iris test harness, tools, context management strategies, benchmarks, and evaluation code are on GitHub. The data construction and training pipelines are planned for later release.