Accio-Lab releases Occamy-1.0 — a 35B open agent-focused model with weights, code and deployment tools

Accio-Lab has publicly released Occamy-1.0, an open-weight 35 billion-parameter AI model aimed at long, tool-using workflows, along with code and documentation for running it. The release is backed by a new paper, “Occamy-1.0: Open Pareto-frontier 35B Intelligence for Co-work,” posted to arXiv as 2609.11977. The arXiv record shows version 1 was submitted Sept. 4, 2026, and the PDF is dated Sept. 14, 2026.

The pitch is practical rather than purely academic. The authors, identified in the PDF as the “Accio Team,” describe Occamy-1.0 as a model for “co-work agents” — systems built to carry out long workflows involving information gathering, tool use, coding and file manipulation across many model calls. In the paper’s abstract, the team says, “We release the model weights and a subset of the training data to support research on practical co-work agents and agentic post-training.”

Occamy-1.0 was not trained from scratch. It is a post-trained specialization of Qwen3.6-35B-A3B, an existing open model, with the new work centered on agent-focused fine-tuning, reinforcement learning and evaluation. According to the published technical details, the model has 35 billion total parameters but activates about 3 billion per token because it uses a mixture-of-experts design, a setup that routes each token through only part of the network. The architecture includes 40 layers, 256 experts, with eight routed experts plus one shared expert active at a time, and a base context length of 262,144 tokens. The model is available on Hugging Face under Accio-Lab/Occamy-1.0, and the supporting code repository is public on GitHub under Accio-Lab/occamy. Both are listed under Apache License 2.0.

In the paper and model card, the authors report results across a range of agent and automation benchmarks, including Claw-Eval, WildClawBench, AutomationBench, Terminal-Bench 2.1, OfficeQA Pro, BFCL v4 and IFEval. They say Occamy-1.0 improves substantially over its starting checkpoint on many agentic tasks and is competitive with some larger frontier systems on several benchmarks. Their headline claim, however, is about efficiency: under the team’s own evaluation and pricing protocol, Occamy-1.0 sits at the “low-cost knee” of a cost-performance Pareto frontier across four representative benchmarks. That comparison uses a pricing snapshot frozen on Sept. 2, 2026, and should be understood as a protocol-specific result described by the authors, not an independently verified universal ranking.

What makes the release notable is that it is immediately usable. Alongside the weights and paper, Accio-Lab has published deployment guidance, including local and private serving options, and released the Dressage training stack used for rollout and reinforcement learning infrastructure. In a year defined by a push toward open, agent-focused models — systems designed to do multi-step work rather than simply answer chat prompts — that combination of open weights, permissive licensing and runnable tooling matters to researchers and developers who want to test or adapt the model themselves.

There is, however, one important transparency caveat. The paper says the team is releasing model weights and “a subset of the training data,” but a single clearly labeled public dataset artifact matching that claim was not obvious in the public model and repository pages reviewed. Related Accio-Lab datasets are public, but the exact Occamy-1.0 training subset was not clearly identifiable from the materials reviewed. The paper’s appendix also notes that some quantities are not fully public, stating: “These quantities are tracked in sealed internal manifests.”