AllSpark releases Iris-mini and Iris-pro, the strongest open-weight search agents in their size class
The Chinese lab AllSpark has released Iris-mini and Iris-pro, two open-weight search agents that now lead their respective size classes on standard benchmarks. The team also published a full training recipe, including the data construction pipeline and reinforcement learning setup, with the goal of letting others reproduce and extend the work.
What makes Iris-mini and Iris-pro different?
Both agents are built on Qwen-series models: Iris-mini uses Qwen3.6-35B-A3B with 35 billion parameters, and Iris-pro uses Qwen3.5-397B-A17B with 397 billion parameters. Both operate with a 256,000-token context window. According to the team’s paper, they deliver the strongest results among open-weight search agents in their size class, with Iris-mini leading on three of four benchmarks and Iris-pro leading or tying in the larger class.
Search agents are language-model systems that research the web on their own: they interpret the question, decide what to search for, read the results, and judge when they have enough evidence to answer. How much of that work the model actually does, versus how much comes from scaffolding around it, is an open question. The AllSpark paper takes that question head on and argues that context management often shifts scores more than the underlying model differences do.
How are training questions generated?
The team reverse-engineers tasks from the link structure of the web. Starting from a seed page and its outgoing links, the pipeline builds a graph of terms and relationships. From that graph it 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 through a simple text search. The agent has to reason across steps rather than look things up directly. Only questions that a reference model cannot solve without tools but can solve with the right sources make it into the dataset. That filter keeps tasks both hard and clearly verifiable.
How is the training data filtered?
A stronger teacher model generates solution paths made up of reasoning, search queries, and results. Each path goes 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 filtering, the model is improved through reinforcement learning against a live web search. The judge model and result summaries run inside the training cluster, powered by 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, where the hardest solved tasks and most efficient solution paths from each round feed back into the next training cycle.
Why does context management matter so much?
During long research sessions the context window can fill up before the agent has resolved every sub-question. Tricks like discarding conversation history extend the research artificially and say little about the model’s actual quality, the team argues. To isolate the effect, they test 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 into what comes from the model and what comes from the scaffolding.
The effect is much larger on the smaller model. Context management boosts Iris-mini’s BrowseComp score by up to 21.2 points. The reason is not a smaller token budget but faster consumption, according to the paper. 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 the 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, then appends that note to the task for the next run.
What did the benchmarks show?
Testing covered four benchmarks: BrowseComp, which tests the ability to find rare facts from indirect clues; BrowseComp-ZH, the Chinese counterpart; DeepSearchQA, which evaluates the completeness of retrieved evidence; and Humanity’s Last Exam, which poses academic questions at expert level. With context management turned on, Iris-mini scored 82.2, 84.8, 86.9, and 52.3 on the four benchmarks according to the paper. Iris-pro reached 88.6, 85.1, 92.9, and 56.4.
In the smaller class, Iris-mini leads on three of four benchmarks and beats the next-best open-weight model on BrowseComp by 3.4 points, though it trails on DeepSearchQA. In the larger class, Iris-pro leads or ties and sometimes approaches systems that require far more compute, according to the authors. All scores come from a single agent run, with no helper agents and no extra verification steps at the end.
What happened with a wrong ground-truth answer?
In the paper’s appendix the team describes a BrowseComp-ZH question about Game of Thrones where the agent was marked wrong even though the answer was backed by the source material. The question targeted Sansa Stark’s second marriage. The agent answered Bolton, while the ground truth said Lannister. Sansa Stark actually marries Ramsay Bolton in her second marriage, so the agent’s answer was correct. The team says contradictions like this between ground truth and source material motivate building better benchmarks.
What else improved beyond search?
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 has been released, and what is coming next?
The model weights for Iris-mini and Iris-pro are available in a collection on Hugging Face, and the code is on GitHub. The release so far includes the Iris Harness with the agent loop, tools, context management strategies, and all four benchmarks with evaluation. The harness runs against any OpenAI-compatible endpoint. 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 two open-weight search agents released by the Chinese lab AllSpark. Iris-mini is built on Qwen3.6-35B-A3B with 35 billion parameters, and Iris-pro is built on Qwen3.5-397B-A17B with 397 billion parameters. Both operate with a 256,000-token context window.
How do Iris-mini and Iris-pro perform on search agent benchmarks?
With context management turned on, 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 on the same benchmarks. Iris-mini leads the smaller size class on three of four benchmarks, and Iris-pro leads or ties in the larger class.
Where can the Iris models and code be downloaded?
The Iris-mini and Iris-pro model weights are available on Hugging Face, and the code is on GitHub. The current release includes the Iris Harness with the agent loop, tools, context management strategies, and all four benchmarks with evaluation, and the team plans to release the data construction and training pipelines later.
Related coverage
- Zhipu AI releases GLM-5.3, claims strongest open-weight coding model with emerging cyber capability
- NVIDIA Expands Open Models for Local AI Agents on RTX and DGX Hardware
This article summarizes reporting from the-decoder.com.