The research paper introduces Occamy-1.0, a co-work agent model developed through staged post-training of the Qwen3.6-35B-A3B checkpoint. The model’s architecture utilizes execution-grounded data and environments, enabling the capture of replayable long-horizon trajectories across multiple harnesses. This approach focuses on consolidating complementary execution capabilities.
Across a suite of co-work benchmarks, Occamy-1.0 consistently ranks among the strongest models of comparable size. Performance is competitive with substantially larger frontier systems on several tasks. The model’s aggregate performance, evaluated across four representative benchmarks, places it at the low-cost knee of the observed cost--performance Pareto frontier.
Supporting evaluations in tool calling, coding, and instruction following demonstrate the model’s ability to preserve broad agentic capability despite specialization. The research team released the model weights and a subset of the training data. This is intended to support research on practical co-work agents and agentic post-training techniques.
The project includes a focus on cost-efficient agent design. The research team aims to provide a model that can be deployed effectively in production environments where cost and latency are key considerations. The release of the model and training data facilitates further experimentation and development in this area.
Source: https://arxiv.org/abs/2609.11977
