Two new open-weight search agents from the AllSpark team now lead their respective size classes on standard web-research benchmarks. Iris-mini (35 billion parameters) and Iris-pro (397 billion parameters), both built on Qwen-series base models, ship with a complete training recipe and the code needed to run them. According to the team’s paper, the same training data and models also lift performance on tasks they were never trained for, including general tool use and office work.
What are Iris-mini and Iris-pro?
It is unclear how much of that work the model itself does versus how much is handled by scaffolding around it. On established benchmarks, leading search-agent systems mostly use the web to confirm knowledge they already absorbed during training.
Iris-mini is based on Qwen3.6-35B-A3B and Iris-pro on Qwen3.5-397B-A17B. Both work with a 256,000-token context window. Across four benchmarks, Iris-mini leads its size class on three of four tests, and Iris-pro leads or ties in the larger class, sometimes approaching systems that need far more compute.
How does the training pipeline work?
The training pipeline builds tasks backward from the link structure of web pages. Starting from a seed page and its outgoing links, it 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 through a simple text search.
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 the tasks both hard and clearly verifiable.
How does filtering work?
A stronger teacher model generates solution paths made up of reasoning, search queries, and results. These 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, 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.
Why does context management matter so much?
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 the conversation history extend the research artificially but 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 into what comes from the model and what comes from the scaffolding around it.
Context management has a much bigger effect on the smaller model, boosting BrowseComp scores by up to 21.2 points. The reason is faster consumption rather than a smaller token budget. 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. That note gets appended to the task for the next run.
What did the benchmarks show?
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 on 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 model, XYZ-Aquila-mini, on BrowseComp by 3.4 points, though it trails on DeepSearchQA. Iris-pro leads or ties in the larger class.
When the ground truth is wrong
In the paper’s appendix, the team describes a case where its 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.” Sansa Stark actually marries Ramsay Bolton in her second marriage. The agent’s answer was correct. The team says contradictions like these between ground truth and source material motivate them to build better benchmarks.
Does training transfer to other tasks?
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 has been released?
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 software with the agent loop, tools, context management strategies, and all four benchmarks with evaluation. The Iris harness can connect to 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 web search agents from the AllSpark team. Iris-mini has 35 billion parameters and is built on Qwen3.6-35B-A3B, while Iris-pro has 397 billion parameters and is built on Qwen3.5-397B-A17B. Both use 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 scores 82.2 on BrowseComp, 84.8 on BrowseComp-ZH, 86.9 on DeepSearchQA, and 52.3 on Humanity’s Last Exam. Iris-pro reaches 88.6, 85.1, 92.9, and 56.4 on the same four benchmarks.
Where can the models and code be downloaded?
The Iris model weights are on Hugging Face, and the code is on GitHub. The data construction and training pipelines are planned for a later release.
Related coverage
- Cisco Antares: Open-Weight AI Models for Vulnerability Localization
- A More Powerful Publicly Available Mythos-Class Model: What It Means for AI Search Visibility
- Poolside Laguna S 2.1 open-weight coding model
SEOScanPRO
SEOScanPRO runs a full technical audit of a site and shows the measured result behind every check. Open the AI visibility report.
Want help with this in Myrtle Beach or Long Island? Talk to us.
NewSunSEO
Grow in Google and AI Search
AI-SEO, local search, paid ads and websites from one team, with offices on Long Island and in Myrtle Beach.
- AI-SEO
- Local AI-SEO
- AI Consulting
- PPC
- Web Design
- Web Development
- Long Island Office(631) 320-1700
- Myrtle Beach Office(843) 492-7131
Request a Free Proposal
Tell us about your business and where you want to grow.
We will reply by email to talk through your goals and what a proposal would include.
We use your details only to reply to you and send what you asked for. See our Privacy Policy.
