Chinese lab AllSpark has released Iris-mini and Iris-pro, two open-weight search agents of 35 billion and 397 billion parameters, together with a full training recipe. According to the paper, the strongest results in their size class come from a single agent with no helper agents and no extra verification steps. For companies choosing between closed search APIs and models they can host themselves, this is the first time a full pipeline — data construction, filtering and reinforcement learning — has been published alongside the weights.
What AllSpark released and how it was trained
The two models build on Qwen-series bases, Qwen3.6-35B-A3B for Iris-mini and Qwen3.5-397B-A17B for Iris-pro, and both work with a 256,000-token context window. 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 and generates a multi-step question whose answer requires chaining several connected steps. 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 enter the dataset, which keeps the tasks hard and verifiable.
Solution paths are generated by a stronger teacher model and then pass 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 improves 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, with the hardest solved tasks and the most efficient paths from each round feeding the next cycle.
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; 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 and sometimes approaches systems that need far more compute, according to the authors.
What this means for companies adopting AI agents
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, and tricks like discarding the conversation history extend the research artificially without saying much about model quality. The effect is uneven: context management boosts the smaller model's BrowseComp score by up to 21.2 points, not because its token budget is smaller but because it consumes tokens faster, needs more steps for the same task and hits the limit more often. For a small company running an agent on a single server, that means the harness around the model may matter more than the parameter count on the label.
The practical consequence is that a 35-billion-parameter agent with well-built context handling can cover a meaningful share of research and tool-use work that previously required a much larger hosted model. The release includes the Iris Harness with the agent loop, tools, context management strategies and all four benchmarks with evaluation, and the harness runs against any OpenAI-compatible endpoint. The weights are available in a collection on Hugging Face and the code is on GitHub, while the data construction and training pipelines are promised for later. A large company can fine-tune on its own document graph; a smaller one can start with the published harness and swap the endpoint.
What the release does not settle is how much of the benchmark result comes from the model and how much from the scaffolding. The team tested every benchmark with and without context management while keeping tools, context limits and the judge model constant, and notes that results reported only with management turned on cannot be cleanly split between the two. The best scores on Humanity's Last Exam come from combining history discarding with a second attempt: if the first try fails, the system condenses it into a short note recording what was already checked and ruled out, and that note is appended to the task for the next run. Buyers should ask vendors which configuration produced the published numbers and whether the same setup is reproducible on their own data.
The clearest marker to watch is the promised release of the data construction and training pipelines. If AllSpark publishes them and outside teams reproduce the reported scores on BrowseComp and DeepSearchQA with their own harness, search stops looking like a narrow specialty and becomes a base skill that transfers to general tool use and office work — the side effect the authors report for both the generated data and the specialized models. If the pipelines stay closed, the weights remain useful but the recipe does not, and the advantage stays with whoever can rebuild the data graph.
