AllSpark Research has launched Iris, an open-source search agent that challenges closed-source competitors. Iris utilizes a Mixture-of-Experts architecture and features a 256K context window, with models available under an Apache 2.0 license on Hugging Face. The agent's effectiveness is attributed to its advanced context management strategies, which prioritize managing search history over simply increasing parameter count, particularly for non-English tasks. Iris was trained using a novel SFT-RL climbing method that incorporates real-time search requests, and its development pipeline is available on GitHub. AI
IMPACT Sets a new benchmark for open-source search agents, potentially accelerating development in tool orchestration and context management.
RANK_REASON Open-source release of a new AI agent with novel architecture and training methods. [lever_c_demoted from frontier_release: ic=1 ai=1.0]
- AllSpark Research
- BrowseComp+
- BrowseComp-ZH
- GitHub
- Hugging Face
- iris
- Iris Search Agent
- mixture of experts
- SFT-RL
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →