An open-weight model built to be an on-device Mac agent.
weights · privateA 9B model, fine-tuned from Qwen3.5-9B, that reads the macOS Accessibility tree and operates apps with one structured tool call per turn. It runs locally, including a 4-bit MLX build.
highlights
- 01Released as a full model, a LoRA adapter and an MLX 4-bit build.
- 02Synthetic trajectories from deterministic simulators of 9 Mac apps (Notes, Finder, Mail, Reminders, Calendar, Settings, Messages, Music, Terminal), with validation, leakage checks and template splits.
- 03MacBench: a 100-episode held-out benchmark plus live-gate, red-team and privacy-probe task sets. The model card reports 97.7% action match.
- 04A reference harness with a deterministic guard that intercepts destructive actions (file deletion, payments, credentials, network) whatever the model says.
evaluation & scope
The reported 97.7% is action match on held-out MacBench tasks for the BF16 model, not end-to-end success across arbitrary Mac workflows. Destructive and external actions require a deterministic runtime guard. Quantized builds should be evaluated separately on the target hardware.
built with
- python
- qwen3.5-9b
- lora
- mlx
- macos accessibility
- hugging face