Researchers trained an open source AI search agent, Harness-1, that outperforms GPT-5.4 on recalling relevant information

Researchers from UIUC, UC Berkeley, and Chroma have developed Harness-1, an open-source AI search agent with 20 billion parameters that significantly outperforms larger proprietary models at retrieving relevant information. The model achieved a 73% accuracy rate on complex retrieval tasks, surpassing GPT-5.4's 70.9% performance and beating other open-source competitors by over 11 percentage points. Harness-1 was evaluated against eight demanding benchmarks that simulated real-world research scenarios, including searches through financial documents, patent databases, and multi-step question-answering tasks requiring logical connections across multiple sources. The model's efficiency stems from its design that separates search management from the core AI system, reducing the "search amnesia" problem where models lose track of information during lengthy retrieval sessions. Notably, the model and its code are freely available under the Apache 2.0 license on Hugging Face, making advanced search capabilities accessible to developers without proprietary restrictions.

This development matters because it demonstrates that smaller, open-source AI models can match or exceed the performance of much larger proprietary systems, potentially democratizing access to sophisticated information retrieval technology for enterprises and developers.