Denny
What must already be on the device
A local AI application needs model weights and a runtime that supports them. Download the model and required components while you have internet. The supported model format and quantisation depend on the application and device.
Storage and memory are different constraints. A model that fits on disk can still exceed the memory available for inference. Start with a supported smaller option and keep the conversation short during the first test.
An offline test you can reproduce
- Finish the model download and open it successfully.
- Disable the network connection.
- Start a new conversation.
- Ask a fresh question whose answer is not already in the history.
- Check that a new reply is generated.
Reading an old message or seeing a cached page does not confirm offline inference. Reconnect if the app needs a missing component, then repeat the test.
Text, voice and search have separate requirements
Local text processing does not automatically make speech recognition, spoken responses or internet search available offline. Inspect the chosen speech components and external services. Search needs a network connection.
Also consider the sensitivity of the material. When using a local model, confirm that the selected route is local rather than an external provider with a similar name.
If the model is slow or will not load
Check the exact format, available storage and free memory. Try a smaller supported model and a shorter context. Device heat, battery settings and other applications can affect performance; a file-size comparison is not a speed benchmark.
Denny includes local AI setup in the application. On current full Denny for Mac, use Apple Silicon M1 or later and macOS 14 or later. For Android availability and compatible models, confirm your build with AFTEAM support.