ClipStart in Settings → AI
The Engine section shows where each task runs. Models shows what is installed and ready to use. For clip suggestions, choose a supported model that accepts images and can answer using your logging fields.
Download the model separately from the app, then try a few clips before starting a long batch. Use the choices offered by your installed build; screenshots may show a different setup.
Choose a model for the work
The current catalog includes these options. Check Settings → AI → Models for availability and setup on your machine.
| Model | Use it for | What to check |
|---|---|---|
| Qwen2.5-VL 7B Instruct, Q4_K_M | General visual logging | Model files installed and a sample request completes |
| Qwen2-VL 2B Instruct, Q4_K_M | Visual logging with less memory | Accuracy on small details and your fields |
| Qwen3.8 27B, Q4_K_M | Larger visual model | Enough memory and time for the batch |
| Gemma 4 26B-A4B Instruct, Q4_0 | Larger visual model | Memory use and support in your model server |
| Whisper Large v3 Turbo | Speech transcription | Language and accuracy on your recordings |
Transcription and visual logging use different models. Installing a speech model does not also install one for images.
Leave room for the edit
The model, ClipLogger and your NLE all use memory. Try a short clip while your usual editing tools are open. If the machine struggles, choose a smaller supported model or run the batch when you are away.
The hardware page estimates about 2–4 minutes per clip on M3 Ultra and 3–6 on M4 Max for a roughly 30B four-bit vision model. Those are modeled examples, not measured promises for your Mac. Time a sample of your own footage before planning the batch.
Use a server you control
For a compatible LM Studio or Ollama setup, enter its server address in the local endpoint settings. Confirm the server is running, the model accepts images and a small request completes. Compatibility depends on the model and server configuration.
Localhost means this Mac. An address on your network sends selected frames and enabled context to that other computer. ClipLogger uses the endpoint you configure; it does not automatically divide a batch across a pool of Macs.
Run the batch when it suits you
Choose Local for your own processing, Your API for an eligible cloud-provider setup, or Rush for hosted processing without model-server configuration. Check the selected route and any cost before starting.
If a local run fails, inspect its error and endpoint status before retrying. Leave an archive batch running overnight.