Occamy-1.0: An Open Cost-Efficient 35B Co-Work Agent Model on the Pareto Frontier

Forum topic · 小凯 · 2026-09-15

Summary

Occamy-1.0 is a cost-efficient co-work agent model built by further post-training the Qwen3.6-35B-A3B checkpoint, introduced in an arXiv paper (2609.11977). Co-work agents execute multi-step workflows combining information gathering, tool use, coding, and file manipulation, where cost and latency accumulate across many model invocations. The authors construct execution-grounded data and environments, capture replayable long-horizon trajectories across multiple harnesses, and apply staged post-training to develop and consolidate complementary execution capabilities such as state tracking, coordination, recovery, and follow-through. On broad co-work benchmarks, Occamy-1.0 ranks among the strongest models of its size and remains competitive with much larger frontier systems on several tasks; under the paper's evaluation and pricing protocol, it sits at the low-cost inflection point of the observed cost-performance Pareto frontier. Supporting evaluations on tool calling, coding, and instruction following show the specialization retains broad agentic ability. Model weights and partial training data are open-sourced.

Overview

The paper "Occamy-1.0: Open Pareto-frontier 35B Intelligence for Co-work" (arXiv: 2609.11977, published 2026-09-15) presents a cost-efficient model for agentic co-work scenarios.

Key Points

Motivation: Co-work agents execute complex workflows combining information gathering, tool use, coding, and file manipulation across many model invocations. Since cost and latency accumulate over the full episode, practical value depends not only on peak capability but on how efficiently capability is delivered. Many everyday work steps emphasize state tracking, coordination, recovery, and follow-through rather than frontier-scale reasoning.
Approach: Occamy-1.0 is obtained by further training the post-trained Qwen3.6-35B-A3B checkpoint. The team builds execution-grounded data and environments, captures replayable long-horizon trajectories across multiple harnesses, and uses staged post-training to develop and consolidate complementary execution capabilities.
Results: On a broad set of co-work benchmarks, Occamy-1.0 consistently ranks among the strongest models of its size and stays competitive with much larger frontier systems on several tasks. Under the paper's prescribed evaluation and pricing protocol, its combined performance on four representative benchmarks places it at the low-cost inflection point of the observed cost-performance Pareto frontier.
Generality: Supporting evaluations on tool calling, coding, and instruction following show that this specialization preserves broad agentic capabilities.
Openness: The model weights and a subset of training data are released to support research on practical co-work agents and agentic post-training.

Reference

arXiv: https://arxiv.org/abs/2609.11977

*Auto-collected on 2026-09-15.*