IBM Granite foundation
IBM Granite provides the open foundation for NotchLive's meeting-specific post-training.
NotchLive is building a meeting-specific model on IBM Granite. It will run entirely on your Mac and work as an optional alternative to Apple Intelligence.
IBM Granite · On-device on Mac · No customer training data
IBM Granite provides the foundation. NotchLive adds meeting-specific post-training, a structured meeting workflow, and fully local processing on the Mac.
IBM Granite provides the open foundation for NotchLive's meeting-specific post-training.
NotchLive adapts Granite with curated historical public meeting records and the structured outcomes people need from meetings.
Meeting text will be processed locally. There is no NotchLive cloud inference service for this model.
Meeting Intelligence works today with Apple Intelligence on supported Macs. The NotchLive model will add another on-device choice rather than replacing the current experience.
The NotchLive model is still in development and is not included in the current release.
NotchLive's post-training and evaluation data comes from curated historical public meeting records that are public-domain or licensed for this use. It is separate from the IBM Granite base model's documented pretraining history.
When you generate a brief, the selected model processes the meeting text locally on your Mac. The meeting stays in your NotchLive library and does not become training or evaluation data.
The meeting transcript is provided only to the selected local model on your Mac.
The generated brief remains part of your local NotchLive meeting record.
Your meeting, edits, and generated brief never become training or evaluation examples.
NotchLive is adapting Granite to follow a changing conversation, separate evidence from interpretation, and produce compact results people can review against the transcript.
Before release, NotchLive will test grounding, names, multilingual text, long meetings, action extraction, and visible uncertainty. Supported Macs and measured performance will be published with the model.
Generated claims should be traceable to the meeting record instead of introducing unsupported detail.
Chinese scripts, CJK text, diacritics, names, and requested output languages should remain intact.
When evidence is weak, the model should preserve useful work and show uncertainty instead of filling gaps.
IBM publishes Granite as an open model family under the Apache 2.0 license. Granite has its own documented pretraining history; the policy on this page covers NotchLive's additional meeting-specific post-training and evaluation data.
No. The NotchLive meeting model is in development. Meeting Intelligence available today uses Apple Intelligence on supported Macs.
No. The model will run entirely on the Mac. NotchLive will not operate a cloud inference service for this model.
No. NotchLive never uses customer meetings, personal notes, corrections, exports, MCP activity, support conversations, or telemetry to train or evaluate the model.
NotchLive uses curated historical public meeting records that are public-domain or licensed for post-training and evaluation. Source provenance and rights are reviewed before inclusion.
No. It will be a selectable alternative alongside Apple Intelligence, including for people whose Macs do not support Apple Intelligence and people who prefer another local model.
Use it today with Apple Intelligence on supported Macs. The NotchLive model is the next fully on-device option.