AllSpark has unveiled Iris-mini and Iris-pro, open-source search agents built on Qwen models that top benchmarks among open-weight competitors in their respective size classes.
The two models demonstrate strong performance on web search tasks while maintaining efficiency. Beyond their primary function, Iris-mini and Iris-pro show unexpected capability gains on untrained tasks, including general tool use and office work applications.
The improvements stem from enhanced training data and model architecture choices that transfer knowledge across domains. This generalization suggests the underlying Qwen foundation provides robust capabilities beyond narrow search optimization.
Both models are available as open-source releases, enabling developers to deploy and modify them for custom applications. The breadth of benchmarked performance—from search to tool use—positions them as viable alternatives to proprietary search agents, particularly for users prioritizing transparency and local deployment.
The release represents progress in open-weight AI agents, addressing the growing demand for capable, accessible models that don't require cloud infrastructure.
Satya Nadella announced a formal Code of Conduct for Microsoft's multimodal AI (MAI) models, emphasizing the company's commitment to deliberate development that prioritizes human control and alignment.
The UK government is piloting three-week AI training bootcamps for 16- to 24-year-olds in Preston, with apprenticeship placements offered to participants upon completion.
Y Combinator's Garry Tan is advocating for US open-weight AI labs to develop distilled versions of frontier models. The push aims to democratize access to advanced AI capabilities while maintaining domestic competitiveness.