Occamy-1.0 ships as a cost-efficient 35B model for co-work agents

Co-work agents chain together many separate calls to a language model to gather information, use tools, write and edit code, and manipulate files over the course of a single task. Because the cost and the latency of every one of those calls add up across the full episode, the authors argue that how much capability an agent delivers matters together with how efficiently it delivers that capability. They also say that many of the individual steps inside everyday work call for state tracking, coordination, recovery from errors, and following a task through to completion, rather than the frontier-scale reasoning that the largest and most expensive models are built around.

Building on that argument, the authors present Occamy-1.0, a co-work model obtained by further training the post-trained Qwen3.6-35B-A3B checkpoint. To train it, they built execution-grounded data and environments, captured replayable long-horizon trajectories across multiple agent harnesses, and applied staged post-training to develop and then consolidate the different execution skills a long agent workflow needs.

The paper reports that across a broad suite of co-work benchmarks, Occamy-1.0 is consistently among the strongest models of comparable size and stays competitive with substantially larger frontier systems on several individual tasks. Under the authors' own evaluation and pricing protocol, its aggregate performance across four representative benchmarks places it at the low-cost knee of the observed cost-performance Pareto frontier. Separate evaluations of tool calling, coding, and instruction following are reported to show that specializing the model for co-work execution does not come at the expense of its broader agentic capability.

The authors release the Occamy-1.0 model weights along with a subset of the training data behind it, saying the goal is to support further research into practical co-work agents and into agentic post-training more broadly.

Key facts

Occamy-1.0 is obtained by further training the 35B-parameter Qwen3.6-35B-A3B checkpoint through staged post-training rather than training a model from scratch.
To build it, the team constructed execution-grounded data and environments and captured replayable long-horizon trajectories across multiple agent harnesses.
Across a broad suite of co-work benchmarks, Occamy-1.0 is consistently among the strongest comparably sized models and stays competitive with substantially larger frontier systems on several tasks.
Under the authors' own evaluation and pricing protocol, aggregate performance across four representative benchmarks places Occamy-1.0 at the low-cost knee of the cost-performance Pareto frontier.
The authors release the Occamy-1.0 model weights and a subset of the training data to support research on co-work agents and agentic post-training.

Why it matters

Co-work agents string together many separate model calls over the course of one task: gathering information, calling tools, writing code, editing files. Because cost and latency accumulate across that whole episode rather than any single call, the authors argue that a co-work agent's real value depends not only on how efficiently it delivers its capability but also on how capable it is at its peak. They add that many of the individual steps in everyday work call for state tracking, coordination, recovery from errors, and following a task through to completion, not the frontier-scale reasoning that the biggest, most expensive models are built for. Occamy-1.0 targets that gap directly: a model tuned for the execution side of long agent workflows.

Who it affects

The model targets teams and researchers building co-work agents: systems that chain together many model calls to gather information, use tools, write code, and edit files over a task. Because the paper's own argument is that cost and latency accumulate across a whole agent episode, teams running such agents at scale are the direct audience for a model that the authors say stays competitive with substantially larger frontier systems on several tasks. Releasing the model weights and part of the training data also targets researchers working on agentic post-training, giving them a documented, cost-efficient reference point alongside the largest, most expensive systems.

How to use it

The authors say they release the Occamy-1.0 model weights along with a subset of the training data used to build it, intending the release to support further research on practical co-work agents and on agentic post-training. The abstract does not give a release date, licence terms, or a download location for either the weights or the data, and it does not say how large the released training-data subset is, so where to obtain them and under what terms is not established here.

How solid is it

This is an arXiv preprint, not a peer-reviewed publication, and its central claims rest on the authors' own evaluation and pricing protocol, run across what the text calls 'a broad suite of co-work benchmarks,' with separate checks on tool calling, coding, and instruction following meant to show the co-work specialization has not narrowed the model's broader agentic ability. The Pareto-frontier placement specifically comes from an aggregate score across four representative benchmarks. The abstract does not name those four benchmarks, does not name the 'substantially larger frontier systems' Occamy-1.0 is compared against, and gives no specific scores, percentages, or point values for Occamy-1.0 or for any comparison model, so neither the size of its advantage nor the identity of what it is measured against can be checked from the text alone.

Risks and caveats

The evaluation and pricing protocol behind the Pareto-frontier claim is explicitly the authors' own, so the result reflects their chosen framework and cost model rather than an independent, third-party benchmark. And since the abstract gives no licence, release date, or hosting location for the released weights and training-data subset, whether and how completely the release reaches outside researchers is not yet established.