Quick Facts
- Harness-1 achieves 73% accuracy on information recall tasks, beating GPT-5.4’s 70.9% and leading open source competitors by 11.4 percentage points
- The 20-billion parameter model uses structured memory management instead of maintaining large context windows, reducing costs while improving performance
- Available immediately under Apache 2.0 license with model weights on Hugging Face
Researchers from the University of Illinois at Urbana-Champaign, UC Berkeley, and Chroma have developed Harness-1, an open source AI search agent that outperforms GPT-5.4 on information recall benchmarks despite being significantly smaller.
Harness-1 scored 73% average accuracy on recalling relevant information from curated datasets, beating GPT-5.4’s 70.9%. The model outperformed the previous leading open source search agent, Tongyi DeepResearch 30B, by 11.4 percentage points. Only Opus-4.6 managed to edge out Harness-1 in overall average performance among frontier models.
The breakthrough comes from a novel architectural approach that moves search session management out of the model’s working memory into a structured software environment. Instead of training the model to handle massive context windows, Harness-1 uses a controlled interface for search operations.
“The harness keeps a working memory with candidate docs, curated evidence, importance tags, search history, evidence links, verification records, dedup/compression, and context-budget markers,” the researchers explained. This approach trains the model to search, curate, revisit, verify, and submit results through structured operations.
The model demonstrated strong performance across multiple domains in testing. Benchmarks included open web searches, SEC financial filings, USPTO patent databases, and multi-hop question-answering tasks requiring logical connections across multiple documents.
Harness-1 maintains frontier-level performance at what researchers describe as “Context-1-level cost and latency.” The budget-aware harness strictly manages context windows rather than allowing continuous expansion, enabling autonomous enterprise deployment without exponential token costs.
The timing proves significant for enterprises needing AI agents that can execute multi-step research across proprietary databases without hallucinating or generating excessive compute bills. The model’s ability to work within structured environments rather than relying on brute-force scaling suggests a new direction for agentic AI development.
The research paper and model code are available under Apache 2.0 licensing, with weights accessible through Hugging Face. This immediate availability gives developers and enterprises access to search capabilities that previously required massive proprietary models.
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.