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

AI summary

Overview

Research area: Agentic AI / large language model post-training for long-horizon digital work automation, with an emphasis on inference cost-performance trade-offs.

Technical level: Advanced. The paper assumes familiarity with supervised fine-tuning (SFT), reinforcement learning for LLMs, MoE architectures, context compaction, and trajectory replay infrastructure.

Scope: A system report describing how a 35B-parameter Mixture-of-Experts model was post-trained, environment-grounded, and evaluated to become a cost-efficient "co-work" agent that sits near the low-cost knee of the observed capability-versus-cost Pareto frontier.

What This Paper Is About

Most digital work agents are judged by peak benchmark scores, but real workflows invoke the model dozens or hundreds of times per task, so cost and latency compound across an episode. Many steps in everyday work require state tracking, tool discipline, recovery, and follow-through rather than frontier-scale reasoning, making it wasteful to route every step through a large model. This paper asks whether a compact 35B model can shift the capability-cost Pareto frontier for sustained, multi-step "co-work," and describes the data pipeline, execution infrastructure, and staged training recipe built to find out.

Key Contributions

1. Execution-grounded data and environments. A traceable pipeline connecting task construction, runnable environments, realized tool behavior, trajectory collection, and task-level grading. Two synthesis routes (environment-first and capability-first) converge on a shared canonicalization, verification, and admission process that checks executable feasibility, semantic alignment, discoverability, and grader discrimination.

2. Harness-aware infrastructure for long episodes. Adapters preserve harness-specific prompt, tool, history-rewrite, and session semantics, while the training layer consumes a single shared trajectory schema. Token-exact capture via an OpenAI-compatible proxy plus a Token-in-Token-Out (TiTO) invariant preserves sampled token IDs, and environment state replay reconstructs files, services, and session state so several policy samples can branch from the same prefix and work state.

3. A specialization-and-consolidation training recipe. A Marathon Expert trained with SFT followed by Hierarchical Decoupled Policy Optimization (HDPO) for sustained long-horizon execution, and a Sprint Expert trained with SFT on a broader distribution of shorter-horizon agentic workloads. The two are combined by parameter-space model merging (a uniform model soup), then refined with Single-Rollout Asynchronous Optimization (SAO) on a broad co-work mixture.

4. Open release. Model weights on Hugging Face, a subset of the training data, and Dressage, an open-source release of the internal agentic RL infrastructure (multi-harness execution, proxy-mediated trajectory capture, sandbox integration, multi-segment conversion for RL).

Main Findings

• Strong general co-work performance at 35B scale. Occamy-1.0 scores 82.20 average on Claw-Eval, above GPT-5.6 Sol (81.8) and DeepSeek V4 Pro (81.7), and below Qwen3.8-Max (83.92), despite being a fraction of the size.

• Repeatability, not lucky runs. Its Claw-Eval pass^3 score is 71.40, above GPT-5.6 Sol (68.9) but below Qwen3.8-Max (73.68) and DeepSeek V4 Pro (74.5). The 10.8-point average-to-pass^3 gap is the smallest among comparably sized models, indicating gains come from consistent execution.

• Competitive long-horizon and open-ended work. WildClawBench 49.16 (above all reported comparably sized models, below GPT-5.6 Sol, Qwen3.8-Max, and GLM-5.2). Business Arena final net worth of $79,868, the best among the models reported in that row. CommerceAgentBench 37.40, GDPval 1128, OfficeQA Pro 48.10, and τ³-Bench (Banking) 37.10.

• Specialization does not collapse broader agentic ability. Terminal-Bench 2.1 59.00, AutomationBench Pass^1 27.60 / Partial 69.10, BFCL v4 65.40, VitaBench 41.75, and IFEval 91.53 all remain competitive with much larger systems.

• Large capability gain at modest cost increase. Relative to its Qwen3.6-35B-A3B starting checkpoint, Occamy delivers a large capability gain with only a modest change in per-task inference cost, placing its aggregate performance across four representative benchmarks at the low-cost knee of the observed cost-performance Pareto frontier.

• Merging is a balancing step, not a free win. The uniform model soup improved AutomationBench and Terminal-Bench 2.1 over both parents and kept WildClawBench between them, but introduced a modest Claw-Eval regression relative to the stronger parent. SLERP produced similar Claw-Eval but lower BFCL v4.

• Continuity is the core technical problem. Because harnesses compact, prune, or rewrite model-visible history, a single episode can span multiple model-visible histories while remaining one continuous task with persistent files, tool effects, and prior decisions. Policy outputs are trained only where they were sampled; inherited or rewritten context and tool observations condition the model but are not training targets.

Methodology in Plain English

The team started from an already post-trained Qwen3.6-35B-A3B checkpoint, which supplies strong language, reasoning, and coding ability, and focused their additional training on turning those abilities into reliable execution.

Data. They built tasks as versioned, executable contracts: a public request, an initial world state, a tool specification, and a private grading contract that is frozen before any evaluated rollout and hidden from the model. Tasks come from two directions: starting from real working environments and constructing tasks those environments can support, or starting from capability dependencies abstracted from real work and finding environments to instantiate them. Every task-environment package passes through shared checks for feasibility, alignment, discoverability, and grader discrimination, using reference executions and negative executions to confirm the grader separates real work from no-ops. Admitted episodes must have their task release, recorded environment effects, and frozen grade joinable under one lineage.

Infrastructure. Every policy call goes through a proxy that records exact input and sampled token IDs, loss masks, and episode/run/turn/segment identifiers. Instead of decoding and re-encoding text between turns, the system stitches stored token fragments together so the model sees exactly the sequence it was trained on. History rewrites (summary replacement, pruning) are made explicit and start a new segment, and the model is never supervised on machine-written summaries. Separately, environment state is reconstructed by snapshot/restore, deterministic fixture rebuild, or session resume, validated by fingerprints and grader preconditions. Episodes are sealed before grading and receive exactly one outcome; any mismatch or lineage gap quarantines the episode rather than producing a policy outcome.

Training. SFT uses about 15.0K trajectories and 403.3M tokens (average 26.9K tokens per trajectory), across general agentic tasks, long-horizon interactive agents, terminal and software engineering, and tool-call grounding, with cross-entropy loss on response tokens only and bin-packed sequences. Two experts are then produced: the Marathon Expert (SFT plus HDPO, with an accuracy-conditioned efficiency reward that compares step counts only among trajectories with the same accuracy, so shorter solutions win within a quality stratum without rewarding sloppiness) and the Sprint Expert (SFT over shorter-horizon agentic work). The experts are merged in parameter space — normalized to a common tensor layout, accumulated in float32, and independently verified — then refined with SAO on a broad co-work mixture, with CompactionRL-style continuation across rewrite boundaries so value estimation does not reset when the context representation changes.

Evaluation. Twelve benchmarks across co-work/office tasks, tool calling, coding, and instruction following, run with complete task sets, pinned harnesses, isolated execution environments, and benchmark-native graders where possible, against five same-scale and four substantially larger baselines.

Why This Matters

The paper argues that for agentic workloads, the right objective is not the highest standalone benchmark score but strong end-to-end execution under realistic cost, latency, and deployment constraints. It provides a concrete demonstration that a mid-size model can match or beat much larger frontier systems on several co-work tasks while remaining feasible to self-host, and it documents the infrastructure required to make long-horizon, multi-harness agent training reproducible at all.

Real-world applications:

• Business operations agents — CRM updates, finance workflows, e-commerce operations, and other stateful multi-step processes where a task may span hundreds of model calls.
• Office and document work — professional deliverables grounded in documents, spreadsheets, and policies, evaluated here via GDPval, OfficeQA Pro, and Business Arena.
• Tool- and API-driven automation — workflow orchestration across structured services, evaluated via AutomationBench, BFCL v4, and VitaBench.
• Self-hosted agent deployment — organizations that need to run agents on their own hardware without paying frontier-model per-call prices across long episodes.

Industry relevance: The cost-per-episode framing directly affects the economics of agent products, where per-call cost and latency accumulate over the full task. The harness-aware infrastructure is also relevant to any team training agents against real execution environments, since it addresses the practical problems of token-exact capture, history rewrites, subagent delegation, and attribution of task outcomes to the correct policy tokens.

Future Directions

1. Beyond the current Pareto knee. The paper positions Occamy as one operating point rather than a finished answer; the open question is how much further the frontier can be moved by better data selection, harness variation, or improved cost-aware training objectives.

2. Merging as a principled operation. The uniform model soup helped on some benchmarks and hurt on others, and SLERP traded Claw-Eval for BFCL v4. Better methods for harmonizing complementary expert behaviors without arbitrary regressions remain open.

3. Closing the gap to frontier systems on the hardest tasks. Occamy remains behind the largest models on WildClawBench, OfficeQA Pro, τ³-Bench, and Terminal-Bench, so the question of which capabilities genuinely require frontier scale is unresolved.

4. Scaling and standardization of co-work evaluation. The report relies on a stated pricing and evaluation protocol with some replicated or modified benchmarks; broader agreement on how to measure cost-performance for long-horizon agents would make cross-paper comparisons more meaningful.

5. Remaining limitations. Section 7 is described as summarizing practical lessons and remaining limitations, indicating the authors themselves treat this as a snapshot of a moving target rather than a settled result.

Target Audience

Researchers and engineers working on agentic post-training, long-horizon RL for language models, and agent infrastructure will get the most from this paper, particularly those interested in token-exact trajectory capture, multi-harness training, and cost-aware objectives. It is also useful for practitioners deciding whether a compact self-hosted model can replace frontier API calls in production agent workloads, and for evaluation researchers interested in cost-performance Pareto framing for multi-step tasks. Readers without a background in LLM post-training will find the results section accessible, but the training and infrastructure sections assume substantial prior knowledge.

Authors’ abstract

Co-work agents execute complex workflows that combine information gathering, tool use, coding, and file manipulation across many model invocations. Because cost and latency accumulate over the full episode, their practical value depends not only on peak capability but also on how efficiently that capability is delivered. Yet many steps in everyday work emphasize state tracking, coordination, recovery, and follow-through rather than frontier-scale reasoning. We present Occamy-1.0, a cost-efficient co-work model obtained by further training the post-trained Qwen3.6-35B-A3B checkpoint. We construct execution-grounded data and environments, capture replayable long-horizon trajectories across multiple harnesses, and use staged post-training to develop and consolidate complementary execution capabilities. Across a broad suite of co-work benchmarks, Occamy-1.0 is consistently among the strongest comparably sized models and remains competitive with substantially larger frontier systems on several tasks. Under our stated 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. Supporting evaluations in tool calling, coding, and instruction following further show that this specialization preserves broad agentic capability. We release the model weights and a subset of the training data to support research on practical co-work agents and agentic post-training.