Lumen runs its AI models on your Mac. This page explains the optional packages, how to check which ones your app has, and how developers add them to a build.
Lumen needs no account or API key. It downloads its models once, from Hugging Face, when you choose Get started, and it never sends work to a remote service when a local tool is missing.
Packages in a Lumen build#
| Part | What it adds | How to get it |
|---|---|---|
| Base build | Keyword text and PDF search, local recording, rough notes and actions, and Markdown export | Build the signed app |
| Local AI bundle | Speech recognition, speech-activity filtering, generated meeting notes, speaker labels within a meeting, and reference text and image retrieval | Prepare it explicitly before you build |
| Search package | OCR, rich documents, multimodal retrieval, an optional code index, and more search controls | Prepare it separately, then rebuild |
A build that includes both packages is a full build.
The local AI bundle uses Whisper large-v3-turbo, Silero VAD, Gemma 4 E2B, and EmbeddingGemma 2, plus local speaker-analysis models. The search package has its own model and runtime inventory. You don't choose a model or provider during onboarding.
Check which packages your app has#
- Open Settings → Local tools.
- Read the readiness state.
Lumen verifies bundled packages at launch. If a package is missing, the features it provides are unavailable, and Lumen doesn't fall back to an online service. A raw weight file on its own isn't a compatible runtime package.
Keyword search and the labeled sample meeting need no models. Searching more formats or preparing a transcript needs the relevant local package.
Prepare a build#
This section is for developers building Lumen from source.
Before you begin#
- The preparation commands download several GB and need the documented toolchains. A base build and ordinary app use make no downloads.
- Read model preparation and native search preparation for requirements, pinned provenance, licenses, and evaluation commands.
Prepare and build#
From the source root:
To add speech, meeting notes, and the reference text and image adapter, prepare the local AI bundle:
bash scripts/prepare-local-models.sh --downloadTo add native intelligent search, prepare the search package:
./scripts/prepare-search.shAfter preparation finishes, assemble and sign the app:
./scripts/build-app.sh release
Limitations#
- Artifact checksums verify a package's identity, not its accuracy or whether an executable is trustworthy.
- Imported adapters must respect explicit paths and the sandbox. Text inside your sources can't authorize extra actions.
- The small offline synthetic tests show that the integration works. They aren't broad quality or hardware certification.
- Speaker labels are estimates within one meeting. Lumen doesn't identify people.
What's next#
- Preview status: see what has been checked and what still needs validation.
- Privacy and permissions: see how local tools run inside the app's boundary.