Occamy-1.0

Alibaba's Accio team argues that most steps in everyday agent work are state tracking, coordination, recovery, and follow-through rather than frontier reasoning, so a co-work model should be judged on cost and latency over a whole episode. Occamy-1.0 is their answer: a 35B-A3B model obtained by further training the post-trained Qwen3.6-35B-A3B checkpoint on execution-grounded data and environments, with replayable long-horizon trajectories captured across several harnesses and staged post-training. The 43-author report positions it on the Pareto frontier of capability against cost for workflow execution.