AllSpark released Iris-mini and Iris-pro on September 13, 2026, saying they are the strongest open-weight search agents in their class. It is the company's first major update to its search agent series since the release of Nano Banana Pro.
AllSpark reported benchmark scores of 82.2 and 88.6 on BrowseComp for Iris-mini and Iris-pro respectively, measured on a test that evaluates the ability to find rare facts from indirect clues. That compares with scores of 78.8 for XYZ-Aquila-mini.
Iris-mini and Iris-pro are built on Qwen-series models and target complex search tasks requiring reasoning and multi-step problem-solving. Availability begins with open-source code and model weights on GitHub and Hugging Face, initially for researchers and developers.
"Training questions are reverse-engineered from the web's link structure," said Jonathan Kemper, the paper's author. The training pipeline constructs tasks backward from the link structure of web pages to generate multi-step questions that require reasoning rather than simple text searches.
The announcement follows the release of Nano Banana Pro. AllSpark framed the significance of Iris-mini and Iris-pro as a step forward in creating search agents that can handle tasks beyond their training data.
AllSpark did not say how the models perform without context management, and raised the open question of whether the results come from the models themselves or the scaffolding around them. The team plans to release the data construction and training pipelines later.
Source: thedecoder
