Evaluated as the on-device voice stack for hour-long offline operation, covering local speech recognition, intent handling, response generation, and speech output. Documentation made the offline boundary clear and supported the recommendation over cloud agents.
- What worked
- Positioning around fully on-device operation and model provisioning matched the offline requirement well.
- What got in the way
- On-device language model scope and tablet shell integration details needed careful scoping for the offline design.
