Qwen3-VL-2B fine-tuned into a computer-use agent that does cell segmentation in a headless labeling app

secemp9 · x · 2026-09-10

Inspired by Astra's computer-use abilities, a developer fine-tuned Qwen3-VL-2B into a specialized computer-use agent: it takes screenshots as input and outputs commands to draw polygons, zoom, and navigate — autonomously performing instance segmentation of cells in a headless labeling app. All navigation and labeling in the demo video is done by Qwen itself.

secemp9's take: the ideal model should be good at both tool call-based vision and native vision, with each generalizing to the other — and even better if the model knows which to use at each step while still obeying explicit user prompts like "only use internal vision" or "use only tools". The route of training specialized computer-use agents clearly works.

Original post →

More from coding & agent

coding & agent channel →