Pick a model and click Load — it runs fully on-device via WebLLM.