AllSpark has released Iris-mini and Iris-pro, two open-weight agents designed to research questions on the web, along with the code used to run and evaluate them. Iris-mini has 35 billion parameters, while Iris-pro has 397 billion; both are based on Qwen models and support context windows of 256,000 tokens.
The team reports that the models lead or tie comparable open-weight systems across BrowseComp, BrowseComp-ZH, DeepSearchQA and Humanity’s Last Exam. Those are results from the authors’ paper rather than independent replications. The release currently includes model weights, the agent loop, tools, context-management methods and four benchmarks. Data construction and training pipelines are promised later.
The training process builds difficult questions backward from links between web pages, removes easy text matches and keeps tasks that require tools. A teacher model creates search paths, which are filtered for correctness and then used in alternating rounds of supervised and reinforcement learning.
The work also highlights a benchmarking problem: runtime scaffolding can matter as much as model quality. Discarding old conversation history raised Iris-mini’s BrowseComp result by as much as 21.2 points because the smaller model took more steps and filled its context sooner. AllSpark therefore reports tests with and without context management. That transparency makes comparisons more useful, but developers still need to separate gains from the model itself from gains supplied by the surrounding agent system.