AllSpark Research has put the weights for two search agents on Hugging Face, under the name Iris. The smaller one, Iris-mini, was post-trained from Qwen3.6-35B-A3B, with 35 billion total parameters and 3 billion activated at inference. The larger one, Iris-pro, is built on Qwen3.5-397B-A17B, with 397 billion total parameters and 17 billion activated. Both use a mixture-of-experts architecture with a 256K context window, and the weights are released under the Apache 2.0 license. The accompanying technical report went up on arXiv on September 3.
What separates a search agent from an ordinary chat model is that it has to decide for itself what to search, whether to keep searching after seeing a result, and when it has gathered enough evidence to stop. For the past two years, that capability has mostly belonged to closed-source products, with little on the open-source side to compare against.
Results on four benchmarks
The paper reports results on four benchmarks: BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE. Iris-mini scores 82.2, 84.8, 86.9, and 52.3; Iris-pro scores 88.6, 85.1, 92.9, and 56.4, with DeepSearchQA measured by F1. Both models post the best results among open-source search agents in their respective parameter classes, and the largest gap between them is on BrowseComp.
One number runs counter to intuition. On the Chinese-language BrowseComp-ZH, the 35-billion-parameter small model scores 84.8 and the 397-billion-parameter large model scores 85.1 — a gap of just 0.3 points. The same pair of models differs by 6.4 points on the English BrowseComp. On the Chinese benchmark, the extra parameters bought almost nothing.
Another comparison in the paper points more directly at what matters. The team tested three inference-time context management strategies: doing nothing; "discard-all," which wipes the accumulated tool-call history and restarts from the original question once context passes a threshold; and "retry," which summarizes discarded leads before continuing. Both models improved with context management turned on, and Iris-mini improved by more than Iris-pro did. In other words, for a smaller model, how well it manages a search history that keeps growing matters more than piling on parameters.
Training data reverse-engineered from hyperlinks
Multi-hop search questions are the hardest to construct. Writing by hand a question that takes three or four pages to answer is prohibitively expensive, and it's easy for a model to shortcut it through string matching — if a proper noun in the question happens to match a page, the model can search once and land straight on the answer.
Iris takes the opposite approach: starting from a seed page picked out of the hyperlink structure of a web corpus, it follows the outbound links to distill an entity graph, builds multi-hop chains on top of that graph, and finally rewrites the question so that no clue can be resolved by literal text matching.
Training itself follows what the paper calls an "SFT-RL climbing" process, alternating supervised fine-tuning with reinforcement learning. The reinforcement-learning stage runs against live search rather than offline snapshots; the reward model used for scoring and the module that summarizes observations are both deployed inside the training cluster. That setup means every training step has to make real network requests — the engineering cost is not small.
Our policy is optimized by RL against live search.
Open source finally has a real contender
Zoomed out, open models have largely caught up with closed ones on general chat and coding; search is the area that has lagged furthest behind. The reason isn't hard to see: a search agent's capability is split evenly between the model itself and the scheduling framework of tools around it, and closed-source products tune both together, while the open-source side has often shipped only the model, not the framework.
This time, Iris lays out the model weights, the context management strategy, and the data construction method all at once, with the training pipeline code marked "coming soon." The GitHub repository credits MiroThinker, Relax, ms-swift, and slime among prior projects, the first two of which were open-source attempts at search agents over the past year.
The report places Iris alongside closed-source models such as GPT-5.5 Pro and Gemini 3.1 Pro in the same BrowseComp table, but the paper's own claim of leading "at the same scale" is scoped to open-source models, and it doesn't spell out the closed-source numbers item by item in the summary. Another thing that isn't disclosed is AllSpark Research's institutional affiliation — public reporting has linked the team to a Chinese content platform, but neither the arXiv page nor the GitHub repository names an institution, and none of the nine authors' personal pages list one either.
For anyone who wants to run their own deep-search setup, 35 billion parameters with 3 billion active is a workable size, and paired with the paper's context management approach, it can run on a single machine. The scope of the Apache 2.0 license also leaves plenty of room for commercial use.
Sources: arXiv paper 2609.04304, the AllSpark-Research/Iris repository, CocoLoop, the Hugging Face model page; benchmark scores and parameter counts for both model tiers follow the paper's tables.