Harness-1
#12university-of-illinois-at-urbana-champaign-uc-berkeley-chroma · v1 · seit 2026-06-01 · 4× · zuletzt 29. Juni 2026
33
Momentum
Harness-1 is a 20-billion-parameter search agent (retrieval subagent) developed by researchers from UIUC, UC Berkeley, and Chroma, built on the openai/gpt-oss-20b base model. It was trained with reinforcement learning inside a stateful search harness that externalizes bookkeeping to the environment, leaving only semantic search decisions to the policy. Across eight retrieval benchmarks, Harness-1 achieves 73% average curated recall, outperforming GPT-5.4 (70.9%) and the next strongest open-source search subagent by +11.4 percentage points. Model weights and harness code are publicly available under the Apache 2.0 license.
Historique du momentum
04.04.03.07.
Fonctionnalités
| Parameter Size (Billions) | 20 |
| Reasoning Capability (AIME Score %) | No AIME score documented. Harness-1 is a retrieval agent; measured benchmark performance: 73% average Curated Recall across 8 retrieval benchmarks (Web, Finance, Patents, Multi-Hop QA) |
| Availability Status | Publicly available (Open Source) – model weights and harness code on HuggingFace & GitHub under Apache 2.0 license |
