Roboflow Inference: run Qwen 3.5 for computer use

This tutorial uses Qwen 3.5 with Roboflow Inference to describe images and control a computer through screenshots. The presenter describes Qwen 3.5 as combining vision, language and coding in one model, with support for tool calls.

The walkthrough starts in Roboflow Workflows, where an image feeds into a Qwen 3.5 block. The presenter selects the 2B model, asks which animal appears in an image, and explains the system prompt and generation limit settings, including a 500-token example. The platform supports 0.8B and 2B models in the recording.

To run models locally, the demonstration uses a local inference server. The presenter says a GPU is likely necessary for practical use and recommends an L4 machine. Roboflow's serverless API support is described as forthcoming at the time of recording.

The computer-use example connects a custom Python script to the inference server. For each requested action, the script captures a screenshot and asks Qwen for a location normalized between 0 and 1,000. It converts that output to screen coordinates and executes a click. The presenter reports that the sequence starts a Roboflow training job. This AI agent demonstration follows separate prompts for each button; it does not establish reliability across arbitrary interfaces or tasks.