Occamy-1.0: Open Pareto-frontier 35B Intelligence for Co-work

Details

Co-work agents execute complex multi-step workflows combining information gathering, tool use, coding, and file manipulation. Their practical value depends not just on peak capability but on cost-efficiency over full episodes — most steps involve state tracking and coordination rather than frontier-scale reasoning. Occamy-1.0 is designed around this observation. Built by further training the Qwen3.6-35B-A3B checkpoint, Occamy-1.0 uses execution-grounded training data and replayable long-horizon trajectories across multiple harnesses. Staged post-training develops complementary execution capabilities. Across a broad suite of co-work benchmarks, it ranks among the strongest at its parameter count and stays competitive with substantially larger frontier models on several tasks. The paper reports a cost-performance Pareto analysis: Occamy-1.0 sits at the low-cost knee of the observed frontier for an aggregate of four representative benchmarks, meaning you get competitive capability at significantly lower inference cost than frontier-scale models. Tool calling, coding, and instruction following evaluations show that co-work specialization does not degrade general agentic capability. Model weights and a subset of training data are released, enabling the research community to study agentic post-training at this scale.