Local Mode runs 100% of speech processing, text processing, and local reasoning model operations on your physical Mac. It requires no network requests, no user account login, and works entirely offline (even in airplane mode).
Under-the-Hood Stack
- Speech-to-Text (STT): Powered by a local Whisper.cpp compilation. It leverages Apple's CoreML framework to compile directly for the Apple Neural Engine (ANE) on M-series Macs, minimizing CPU drain.
- Local Reasoning (LLM): Runs local quantized weights for Qwen-2.5-Instruct (or Google Gemma-2B) natively on your CPU/GPU cores.
Bootstrap Initialization
On the first launch in Local Mode, the application backend will trigger a bootstrap script to download model weights. This is a one-time operation.
Downloading Model Weights
Ensure you have a high-speed internet connection during the first setup. Escribolt will fetch approximately 1.8GB of weights. Once download checks pass, the network link is disengaged.
Limitations of Local Mode
While providing total data isolation, please keep in mind:
- Intel-based Macs will experience higher latency and processing delay during transcript generation.
- Cross-device synchronization is disabled, as notes are stored locally in a local SQLite file.