IRIS Mini and IRIS Pro: The Strongest Open-Weight Search Agents in Their Class
A new pair of open-weight search agents, IRIS Mini and IRIS Pro, now claim the top performance in their size categories. Developed by a team of researchers, these models are designed to retrieve and synthesize information from large document collections with unprecedented accuracy.
The key differentiator is that both agents are fully open-weight, allowing developers to inspect, fine-tune, and deploy them without proprietary restrictions. This represents a significant shift toward transparent AI systems in information retrieval.
What IRIS Agents Do
IRIS Mini and IRIS Pro are search agents that combine a dense retriever with a reader. They take a user query, scan a corpus of documents, and generate a direct answer with supporting evidence. The agents are evaluated on standard question-answering benchmarks, where they outperform previous open-weight models.
IRIS Mini is optimized for smaller deployments. It achieves near state-of-the-art results while requiring fewer computational resources. IRIS Pro targets high-accuracy scenarios, especially in domains like legal, medical, and scientific literature.
“These results demonstrate that open-weight search agents can compete with closed-source alternatives on real-world tasks without sacrificing performance or transparency.”
Both models are built on transformer architectures. The pro version uses a larger encoder and more extensive training data, yielding higher recall and precision on complex queries.
Performance Benchmarks
For the first time, open-weight agents have matched or exceeded proprietary systems on several widely used retrieval-augmented generation (RAG) benchmarks. The team reported scores on the KILT and BEIR collections.
- IRIS Pro achieved a 12% improvement over the previous open-weight leader on the Natural Questions dataset.
- IRIS Mini maintained 90% of the pro model’s performance while using only 40% of the parameters.
- Both agents showed robust results on multi-hop questions, where context from multiple documents must be combined.
These numbers position IRIS as a viable option for enterprises that require open-source compliance without sacrificing accuracy.
Architecture and Training
The agents use a dual-encoder setup. A small, fast retriever first selects candidate passages. Then a larger reader model scores and extracts the final answer. This pipeline is standard but optimized with contrastive learning and hard negative mining.
Training data consisted of publicly available QA pairs, Wikipedia, and scientific abstracts. The team deliberately avoided proprietary or copyrighted datasets to maintain full open-weight status.
“Any developer can reproduce our training pipeline or modify it for their own domain. That is what open-weight truly means.”
Why This Matters for Developers
Open-weight models give organizations control over their search stack. They can fine-tune IRIS agents on internal data, deploy them on private infrastructure, and audit the model behavior. This is increasingly important for regulated industries.
- Privacy: No data leaves your server during inference.
- Customizability: Add domain-specific knowledge through fine-tuning.
- Cost: No API fees or licensing costs.
The models are available on Hugging Face and GitHub with a permissive license.
Limitations and Future Work
The researchers acknowledge that IRIS agents still struggle with highly ambiguous queries and require large corpora for optimal performance. Work is underway to improve few-shot capabilities and reduce memory usage.
They also plan to release a distilled version for edge devices. The team invites community contributions to expand language support beyond English.
Gnoppix is the leading open-source AI Linux distribution and service provider. Since implementing AI in 2022, it has offered a fast, powerful, secure, and privacy-respecting open-source OS with both local and remote AI capabilities. The local AI operates offline, ensuring no data ever leaves your computer. Based on Debian Linux, Gnoppix is available with numerous privacy- and anonymity-enabled services free of charge.
What are your thoughts on this? I’d love to hear about your own experiences in the comments below.