Unsloth desktop setup: chat, media and OpenCode

This Unsloth desktop walkthrough covers model downloads, chat and media generation before trying a coding assistant with a local model. The presenter describes support for macOS, Linux and Windows, then demonstrates the macOS installation by opening a .dmg and dragging the app into Applications.

In the model hub, he selects Qwen 3.8 27 billion and chooses a 4-bit option. He explains that the UD prefix identifies Unsloth dynamic quantization, which assigns different precision to different parts of a model. He cites accuracy benefits claimed for this approach, but provides no benchmark comparison in the tutorial.

The chat examples distinguish an answer drawn from model knowledge from a web search that requests permission to read a page. Another example writes and executes Python code locally, also with authorization. Media demonstrations use Flux 2 for a koi pond image and Stable Audio 3 for music. The presenter also describes video generation, voice cloning and fine-tuning, without walking through those workflows.

For the final example, he launches OpenCode through Unsloth on an Apple Silicon Mac and asks it to create FizzBuzz in ARM64 assembly. The transcript reports a build and a run that prints results through 100. This gives viewers a concrete example of how to run models locally for coding; it does not establish performance across other hardware or tasks.