Accio-Lab/occamy

What it solves

Occamy-1.0 is a compact AI model built for real-world collaborative work: tasks that span many steps, involve tools, files, search, and structured APIs, and need to stay coherent over long sessions. It focuses on reliable execution, keeping track of state, recovering from errors, and following through—rather than just answering questions.

How it works

The model starts from an existing checkpoint (Qwen3.6-35B-A3B) and is further trained specifically for agentic co-work. It uses a Mixture-of-Experts design with 35B total parameters but only 3B active per call, making it efficient for long-running workloads. Training involved two specialists—one for sustained execution, one for preserving general ability—then merged, and refined. All training data comes from executable environments with observable outcomes. The model supports long context (up to 262,144 tokens) and can be served with standard tools like SGLang and vLLM.

Who it’s for

Developers and researchers building agentic systems that need a model that can handle multi-step professional workflows, use tools, write code, and interact with files and APIs. It's also for those who want an open model with a compact inference footprint.

Highlights

• Specialized for co-work: long-horizon, stateful tasks across tools, files, and APIs.
• Compact inference: 35B total, 3B active parameters, efficient for extended runs.
• Strong performance on co-work benchmarks, beating its base model and several comparable models.
• Open training stack (Dressage) and multiple quantized checkpoints available.