Understand model recommendations
Outcome
Section titled “Outcome”You understand what the recommendation changes and how to choose a different model when it does not fit your use case.
How recommendations work
Section titled “How recommendations work”At runtime, Synthezia reads available memory, CPU-core count, and processor architecture. It uses those values to suggest a Whisper transcription model and an Ollama summary model. If detection is unavailable, it falls back to smaller, safer defaults.
| Capability | Status | Current behavior |
|---|---|---|
| Machine-based model recommendations | Beta | Recommend Whisper and Ollama model defaults from detected memory, CPU count, and architecture. |
| Whisper Small | Available | Whisper model recommended by the current heuristic for many lower-memory machines. |
| Whisper Large v3 Turbo | Available | Default transcription model and the current Apple Silicon recommendation for many machines with at least 16 GB of memory. |
| Llama 3.2 3B | Available | Lightweight local model exposed by the curated Ollama catalog and recommendation heuristic. |
| Llama 3.1 8B | Available | Default local summary and chat model and the balanced recommendation for higher-memory machines. |
Use the recommendation safely
Section titled “Use the recommendation safely”- Accept it for the first local test if you do not have a reason to choose another model.
- Download only the models you need.
- Test a representative recording and summary on your own Mac.
- Change the default in Models if a different installed model better serves your workflow.
Important limits
Section titled “Important limits”A recommendation is a current heuristic. It does not prove system compatibility, exact download size, required disk space, speed, memory use, transcription quality, or summary quality. Keep a source recording and review all generated results.