ICANEWS

Occamy-1.0: Cost-Efficient 35B Intelligence for Complex Co-work Workflows

arXiv CS · · 2 min read · Engineering & Technology

Read research and analysis on Occamy-1.0: Cost-Efficient 35B Intelligence for Complex Co-work Workflows published by ICANEWS, a global research journal for emerging researchers.

Key Takeaways

  • Occamy-1.0 is consistently among the strongest comparably sized models on co-work benchmarks.
  • The model remains competitive with substantially larger frontier systems on several tasks.
  • Occamy-1.0 is positioned at the low-cost knee of the observed cost-performance Pareto frontier across four benchmarks.
  • Its specialization preserves broad agentic capability in tool calling, coding, and instruction following.

Why This Matters

The model's efficiency in co-work tasks, particularly its low-cost operation relative to its capabilities, addresses a key practical challenge for agents executing complex, multi-invocation workflows. This could broaden the applicability of agentic systems by making their operation more economically feasible.

Overview

Occamy-1.0 represents a cost-efficient intelligence model engineered for complex co-work workflows. It is derived from the post-trained Qwen3.6-35B-A3B checkpoint. The development focused on addressing the cumulative cost and latency associated with multi-invocation co-work agents, which execute tasks combining information gathering, tool use, coding, and file manipulation. The model's design prioritizes both peak capability and the efficiency of its delivery, recognizing that many routine work steps emphasize state tracking, coordination, recovery, and follow-through over frontier-scale reasoning.

Research Context

Co-work agents engage in complex workflows requiring multiple model invocations. The practical utility of these agents is influenced by their cumulative cost and latency across an entire operational episode. While peak capability is a factor, the efficiency with which this capability is delivered is also critical. Daily work often involves tasks centered on state tracking, coordination, recovery, and follow-through, rather than solely relying on the most advanced reasoning capabilities. This context highlights the need for models that balance capability with operational efficiency for practical applications.

Approach

The development of Occamy-1.0 involved further training the Qwen3.6-35B-A3B checkpoint. The methodology included constructing execution-grounded data and environments. Researchers captured replayable long-horizon trajectories across multiple harnesses. A staged post-training process was then utilized to develop and consolidate complementary execution capabilities within the model. This approach aimed to optimize the model for efficiency in co-work tasks.

Findings

  • Occamy-1.0 consistently ranks among the strongest comparably sized models across a broad suite of co-work benchmarks.
  • The model demonstrates competitiveness with substantially larger frontier systems on several tasks.
  • Under the stated evaluation and pricing protocol, Occamy-1.0's aggregate performance across four representative benchmarks places it at the low-cost knee of the observed cost-performance Pareto frontier.
  • Supporting evaluations in tool calling, coding, and instruction following indicate that its specialization preserves broad agentic capability.

Why This Matters

The efficiency demonstrated by Occamy-1.0 in managing co-work agent workflows can be significant given the cumulative cost and latency of multi-invocation tasks. By offering competitive performance at a lower cost, it addresses a practical constraint for agents performing complex, multi-step operations. This could influence the viability and accessibility of agentic systems for everyday work scenarios that require state tracking, coordination, and follow-through.

Potential Applications

The model weights and a subset of the training data for Occamy-1.0 are being released. This release aims to support ongoing research into practical co-work agents and agentic post-training methodologies.

Research Information

Institution
arXiv CS
Original Study
View Publication
Source
arXiv CS

About ICANEWS

ICANEWS is a global research journal for emerging researchers, publishing student and emerging researcher work across all fields.