Iris-mini and Iris-pro Are Now the Strongest Open-Weight Search Agents on Their Benchmarks

The open-weight AI ecosystem has produced strong generalist models and strong reasoning models. Strong search agents — models specifically designed to perform agentic internet research tasks — have been a weaker category, with most competitive options sitting behind closed APIs. Iris-mini and Iris-pro change the availability picture.

What Iris Is

Iris-mini and Iris-pro are open-source search agents developed by the AllSpark team, built on Qwen base models. They’re designed for agentic internet search: the ability to break down a research question, issue multiple queries, synthesize results across sources, and return structured, cited responses.

Both models lead their respective size class benchmarks among open-weight alternatives — meaning the best open-source options currently available for search agent use cases.

The Surprising Finding

What makes the Iris release more interesting than a benchmark leaderboard update is a training result the AllSpark team didn’t plan for. The training data and methods used to develop search capabilities also improved performance on tasks the models were never explicitly trained for: general tool use and office work tasks.

This cross-task generalization is a pattern that researchers have been documenting in language model training more broadly — improvements in one capability domain often produce capability gains in adjacent domains, through mechanisms that aren’t fully understood. In Iris’s case, the practical implication is that these models aren’t specialized search-only tools but appear to function as capable general agents, with search as a first-class skill.

Why Open-Weight Matters Here

Search agents that can access the internet and synthesize information from it are a category with meaningful enterprise deployment sensitivity. A search agent operating on proprietary research questions, competitive intelligence tasks, or regulated information sources benefits from the auditability and deployability that open weights provide. Organizations that cannot send data to a closed API for compliance or security reasons now have a competitive open-weight option that was not previously available at this benchmark level.

The combination of Qwen base models — well-tested, widely deployed, available in multiple size configurations — with AllSpark’s search-specific training gives Iris a deployment profile that closed alternatives can’t match for on-premises or air-gapped use cases.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.