Harness-1: The AI Search Agent Outperforming GPT-5.4

Harness-1: A Game Changer in AI Search

In a remarkable collaboration between the University of Illinois at Urbana-Champaign, UC Berkeley, and Chroma, researchers have unveiled Harness-1, an open-source AI search agent that boasts 20 billion parameters. This model has achieved a stunning 73% accuracy in recalling relevant information, surpassing GPT-5.4's 70.9% and outpacing its closest competitor, Tongyi DeepResearch 30B, by a significant margin.

Harness-1's performance is particularly impressive given its ability to tackle complex search benchmarks across various domains, including:

• Open web searches
• SEC financial filings
• Technical patent databases
• Multi-hop question-answering tasks

By offloading the intricate bookkeeping of search sessions, Harness-1 enhances its efficiency and effectiveness, proving that smaller models can compete with larger, proprietary systems. With its immediate availability under the Apache 2.0 license, developers can leverage this powerful tool to transform their data retrieval processes.