Occamy-1.0 Targets the Cost-Performance Sweet Spot for 35B Co-work Agents

Introduction

A useful work agent must do more than produce a strong answer in a single turn. It may need to gather information, call tools, write code, manipulate files, track intermediate state, and recover when something goes wrong. Because every additional model invocation adds latency and expense, practical agent quality depends on the cost and reliability of the entire workflow, not just peak reasoning performance.

Occamy-1.0, described in a new arXiv paper, is a 35B-scale model built around that premise. The researchers further trained the post-trained Qwen3.6-35B-A3B checkpoint with an emphasis on execution-oriented behavior. Rather than relying only on a larger model, the project asks whether a moderately sized open model can be made more effective for sustained, tool-mediated work.

Key points

• Training is centered on execution. The team created execution-grounded data and environments, then captured replayable long-horizon trajectories across multiple harnesses. This gives training a view of intermediate actions, tool interactions, and recovery behavior, rather than evaluating only the final response.
• Capabilities are developed in stages. According to the paper, staged post-training is used to build and consolidate complementary abilities. These include state tracking, coordination, recovery, and follow-through—features that are essential in everyday workflows but less visible in short benchmark prompts.
• The objective includes full-episode economics. Across a broad set of co-work benchmarks, Occamy-1.0 is reported to be among the strongest models of comparable size. Under the authors’ stated evaluation and pricing protocol, its aggregate result on four representative benchmarks sits near the low-cost knee of the observed cost-performance Pareto frontier. On several tasks, it remains competitive with substantially larger frontier systems.
• The specialization is not limited to one tool scenario. Supporting evaluations cover tool calling, coding, and instruction following. The reported purpose of these tests is to show that execution-focused training preserves broad agentic capability.
• The release supports further research. The authors say they will release model weights and a subset of the training data, enabling additional work on practical co-work agents and agentic post-training.

Why it matters—and what remains unclear

Occamy-1.0 reflects a broader shift in how agent models are judged. The relevant question is increasingly not simply which model is most capable, but which model can complete a long sequence of actions reliably at an acceptable cost. For automation systems that make many calls, state management, tool use, error recovery, and task completion may affect deployment value more directly than a single difficult reasoning step.

The paper therefore makes a case for specialized open models as a practical middle ground between small general models and much larger frontier systems. Still, the supplied material does not include individual benchmark scores, latency figures, prices, or a task-by-task comparison with competing models. The “low-cost knee” should consequently be read as a result under the paper’s stated protocol, not as a universal deployment guarantee. Open weights and partial data release also do not by themselves reveal the full training mixture or real-world stability. Independent reproduction and testing on production toolchains will be important next steps.

arXiv