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

Occamy-1.0. Open Pareto-frontier 35B Intelligence for Co-work. By Wenhui Chen and forty-two others, from the Accio Team. Published on the fourth of September. Its code repository is public but slight: nine source files.

Co-work agents run complex workflows that mix information gathering, tool use, coding and file manipulation, across many model invocations. Cost and latency accumulate over the whole episode, so practical value depends not only on peak capability but on how efficiently that capability is delivered. And many steps in everyday work call less for frontier-scale reasoning than for state tracking, coordination, recovery and follow-through.

Occamy-1.0 is a co-work model built for cost efficiency. It is obtained by further training the post-trained Qwen3.6-35B-A3B checkpoint, specialising an existing model for this kind of work rather than reaching for a larger one. The reasoning follows from the problem. If an episode involves many calls, and most of those calls need reliable execution more than peak reasoning, then a model tuned for execution can deliver that value more cheaply over the whole episode.

Training rests on three pieces. Execution-grounded data and environments are constructed for the purpose. Replayable long-horizon trajectories are captured across multiple harnesses. And staged post-training develops, then consolidates, complementary execution capabilities. Every piece points the same way: data taken from work that was actually carried out, trajectories long enough to include the tracking and recovery that everyday work demands, and training that builds several complementary skills before consolidating them into one model.

Across a broad suite of co-work benchmarks, Occamy-1.0 is consistently among the strongest models of comparable size, and remains competitive with substantially larger frontier systems on several tasks. Under the authors' stated evaluation and pricing protocol, its aggregate performance across four representative benchmarks places it at the low-cost knee of the observed Pareto frontier of cost against performance. The knee is the point where moving to a costlier model starts buying noticeably less performance for each extra unit spent.

Supporting evaluations in tool calling, coding and instruction following show that the specialisation preserves broad agentic capability rather than trading it away.

The model weights and a subset of the training data are released, to support research on practical co-work agents and on agentic post-training. The argument is about fit rather than size. When much of an episode is tracking, coordinating and recovering, a thirty-five billion parameter model trained on execution can sit at the cheap end of the frontier without giving up its broader agentic skills.