# Iris-mini and Iris-pro Set New Open-Source Search Agent Standards
AllSpark has released two open-source search agents, Iris-mini and Iris-pro, built on Qwen models that outperform competing open-weight systems in their respective size classes. The models lead benchmarks for web search and information retrieval tasks while maintaining the flexibility of open-source deployment.
The Iris models represent a shift in how search agents are built and evaluated. Rather than relying on proprietary APIs or closed commercial systems, these agents run on openly available weights. Iris-mini targets resource-constrained environments, while Iris-pro aims at maximum performance without architectural constraints. Both models show measurable advantages on standard benchmarks used to evaluate search agent capabilities.
What makes these releases noteworthy extends beyond raw benchmark numbers. The training data and underlying model improvements generalize to tasks the Iris models never encountered during training. This transfer learning effect means better performance on general tool use and office work applications. That generalization matters because it suggests the training approach captured useful capabilities that extend far beyond search itself.
The use of Qwen models as the foundation provides additional context. Qwen, developed by Alibaba, has become a competitive base for fine-tuning and specialized applications. Building search agents on Qwen rather than other alternatives like Llama reflects the breadth of model choices now available to developers building agent systems.
Open-weight search agents address a real deployment gap. Many organizations cannot rely on external APIs for search queries due to privacy, cost, or latency concerns. Running a search agent locally means faster inference, no data leaving internal systems, and predictable pricing. The tradeoff involves handling indexing and document retrieval infrastructure yourself, which Iris models enable through open deployment.
The benchmark leadership matters for adoption decisions. When organizations evaluate whether to build custom search agents versus deploying existing models, performance metrics drive choices. Iris-mini and Iris-pro now occupy attractive positions in the tradeoff space: smaller models that match larger competitors while smaller models that still deliver strong results.
The generalization to untrained tasks reveals something about how these agents learned to approach problems. Rather than simply memorizing search patterns, the models developed broader reasoning capabilities that apply to tool use and office workflows. This suggests the training methodology emphasized compositional understanding over task-specific optimization.
Open-source release creates network effects around these models. As developers build on Iris-mini and Iris-pro, they contribute improvements and use cases back to the community. This contrasts with proprietary systems where improvements remain locked within single organizations. The open approach has historically accelerated adoption of foundation models like Llama and Mistral.
The search agent space continues fragmenting into specialized systems. Rather than one dominant approach, different models now compete on specific dimensions: speed, accuracy, reasoning capability, and deployment flexibility. Iris-mini and Iris-pro claim leadership in open-weight categories while other systems compete on other axes like reasoning depth or multimodal capability.
For organizations building AI systems, Iris-mini and Iris-pro now represent viable alternatives to closed search APIs. The benchmark results provide confidence that open-weight models can deliver competitive results. Developers can now choose between convenience of managed APIs and control of open deployment, with performance no longer clearly favoring closed systems in these size categories.