A new pair of open-weight search agents from the AllSpark team now posts the strongest results in their respective size classes on four common web-research benchmarks, with both models and a full training pipeline released to the public. The release covers the agents themselves, the training recipe, and the evaluation harness, giving outside teams a way to reproduce the numbers.
Search agents built on language models have to read a question, plan what to look up, read what comes back, and judge when they have enough evidence to answer. How well today’s leading systems actually do that is contested: on standard benchmarks, many of them mainly use the web to confirm knowledge they already absorbed during training, rather than discovering new information.
What AllSpark released
AllSpark, a Chinese AI lab, has published Iris-mini and Iris-pro, two open-source search agents of different sizes, along with the training recipe behind them. The smaller model, Iris-mini, has 35 billion parameters, and the larger one, Iris-pro, has 397 billion. Both are built on Qwen-series models, Qwen3.6-35B-A3B and Qwen3.5-397B-A17B respectively, and both work with a 256,000-token context window.
The team reports that both models set the strongest results among open-weight search agents in their size class. The model weights are available in a Hugging Face collection, 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 used for evaluation. The harness runs against any OpenAI-compatible endpoint, and the team plans to release the data construction and training pipelines later.
How the training data is built
The training pipeline generates tasks by working backward from the link structure of the web. Starting from a seed page and its outgoing links, it 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 the clues cannot be resolved through a simple text search. The agent has to reason across the chain, not just look things up.
Only questions that a reference model cannot solve without tools but can solve once it has the right sources make it into the dataset. That filter 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 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. The hardest solved tasks and the most efficient solution paths from each round feed back into the next training cycle.
Context management changes the numbers more than the models do
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 all sub-questions. Tricks like discarding conversation history extend the research artificially but say little about the model’s actual quality.
To isolate that 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 into what comes from the model and what comes from the scaffolding around it. The Iris scores also come from a single agent, with no helper agents and no extra verification steps at the end.
Benchmark results
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 turned 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. In the smaller class, Iris-mini leads on three of four benchmarks and beats the next-best open-weight model, XYZ-Aquila-mini, on BrowseComp by 3.4 points, though it trails on DeepSearchQA. Iris-pro leads or ties in the larger class and, according to the authors, sometimes approaches systems that require far more compute.
Context management has a much bigger effect on the smaller model, boosting BrowseComp scores by up to 21.2 points. The reason 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 run 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.
A wrong ground-truth answer
In the paper’s appendix, the team describes a case where the agent was marked wrong even though the answer was backed by 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, so the agent’s answer was correct. The team says contradictions like these between ground truth and source material motivate them to build better benchmarks.
An unexpected side effect on other tasks
Beyond search, the authors report that 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.
FAQ
What are Iris-mini and Iris-pro?
They are two open-weight search agents released by the AllSpark team, built on Qwen-series models with 35 billion and 397 billion parameters respectively. Both run with a 256,000-token context window and post the strongest results among open-weight search agents in their size class on four benchmarks.
How are the Iris training questions generated?
Training tasks are reverse-engineered from web link structure. Starting from a seed page and its outgoing links, the pipeline 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 replaced with a paraphrase, so clues cannot be resolved through a simple text search.
What benchmarks were the Iris models tested on?
Testing covered BrowseComp, BrowseComp-ZH, DeepSearchQA, and Humanity’s Last Exam. With context management on, Iris-mini scored 82.2, 84.8, 86.9, and 52.3, and Iris-pro reached 88.6, 85.1, 92.9, and 56.4 across the same four tests.
Related coverage
- Tech Firms Urge Washington Not to Restrict Open-Weight AI
- Poolside Laguna S 2.1 open-weight coding model released
- Cisco Antares: Open-Weight AI for Vulnerability Localization
This article summarizes reporting from the-decoder.com. See our editorial disclaimer for how our articles are produced.
Run a free scan to see your AI Visibility Score, SEO rating, and local citation accuracy.