Local AI on M‑chips
Two recent posts highlight offline AI moving to Apple Silicon: Pewbeam added offline operation on M‑chip MacBooks, and the Tome app is running Parakeet‑TDT v3 transcription locally and piping results into Obsidian vaults. Both underscore growing zero‑cloud workflows for privacy‑focused transcription and assistant features on Macs. (x.com) (x.com)
Pewbeam’s docs now list an on-device macOS transcription path that runs the sermon-to-verse pipeline in “under a second” and performs semantic search across "31,000+ verses." (pewbeam.com)) The Pewbeam project is led by product designer Dára Sobaloju and is presented as live church presentation software that syncs detected scripture to projection outputs. (darasoba.com)) The Tome macOS app’s public repo advertises local meeting and voice‑memo capture that transcribes with Parakeet‑TDT (now referenced as v3 in the project blurb) and writes structured.md files directly into an Obsidian vault. (github.com)) NVIDIA’s Parakeet‑TDT‑0.6B‑v3 is a 600‑million‑parameter multilingual ASR model covering ~25 European languages and is reported on the Hugging Face model card and arXiv release notes. (huggingface.co)) Third‑party tooling and community builds show Parakeet‑TDT v3 running with ONNX/quantized runtimes and CPU‑first implementations that aim for low‑latency transcription on consumer machines, with tutorials and fast‑inference repos emerging in February–March 2026. (github.com)) Model distribution and size are nontrivial for local workflows: community model mirrors list Parakeet‑TDT‑0.6B‑v3 downloads at roughly 2.5 GB and advertise native punctuation, capitalization, and word‑level timestamps that apps can surface without cloud post‑processing. (modelscope.cn))
