Download for MacA note about RAM, the cloud, and what we actually send.
We want to be straight with you about how Hapi works, what stays on your Mac, and where the realistic trade-offs live. No marketing absolutes — just what we've learned shipping this to real users.
Why we recommend at least 24 GB of RAM
Hapi was built on a simple principle: nothing has to leave your Mac. Recording, transcription, diarization, search, and the language model that powers summaries and chat are all designed to run on-device.
After many rounds of real-world testing, here is the honest picture of what you can expect at each memory tier:
16 GB
Hybrid recommended
Recording and transcription run fine locally, but the LLM cannot fit alongside diarization without memory pressure. The smallest model can run end-to-end, but you should expect summaries and chat quality to be noticeably below a frontier cloud model — fine for short voice notes, less reliable on dense 90-minute meetings. We suggest hybrid and routing AI steps to the cloud.
24 GB+
Comfort floor — 100% local
Best tier (4B) on 24–31 GB, Pro tier (8B) on 32 GB and above. This is the point where local matches cloud quality on most meetings, and the only configuration where we recommend running 100% on-device by default.
The pattern is straightforward: smaller models fit in less memory but produce shorter, less faithful summaries and weaker reasoning in chat. We could pretend otherwise — but you would notice the first time you tried to summarize a long call on a base-tier MacBook Air. So below 24 GB, hybrid is what we genuinely recommend.
Hybrid mode — local recording, optional cloud AI
So we offer a hybrid mode. Recording and transcription stay 100% local on every Mac, and you can optionally route only the AI steps to a cloud model through OpenRouter using your own API key. The steps you can route to the cloud are:
- Chat (“Ask Hapi”)
- Meeting summaries
- Action item extraction
- Pre-meeting briefings
- Entity recognition (people, companies, projects)
- Speaker label suggestions
What never leaves your Mac
Even in hybrid mode, these never touch the network:
- Raw audio recordings. Audio is never uploaded.
- Speech-to-text. Parakeet runs on-device.
- Voice embeddings used for diarization (ECAPA, WeSpeaker, Pyannote).
- Text embeddings used for semantic search.
- The SQLite database with your transcripts, entities, and history.
When text does leave the device in hybrid mode, only the cleaned transcript snippets or summary chunks needed for that specific step are sent — and only to the provider you authorized.
Two layers of protection on every cloud call
Hybrid mode is opt-in. When it's on, we apply two layers on top of standard transport security so that the data we do send is treated as ephemeral, not training material:
Layer 1
Zero-retention provider allowlist
Hapi only dispatches to models whose providers contractually commit to no logging and no training on your inputs. Models without that commitment are filtered out of the picker entirely. This is on by default and you can verify it in Settings → AI.
Layer 2
Per-request opt-out signal
Every cloud call carries an explicit data_collection: deny flag, instructing the provider to discard the prompt and response after generation rather than keeping them for analytics or fine-tuning.
On top of those two layers:
- Traffic is encrypted end-to-end with TLS.
- Your API key is stored in the macOS Keychain — never in plain settings, never proxied through us. Hapi talks to OpenRouter directly from your Mac.
- You can set a monthly budget; if it's hit, Hapi automatically falls back to the local model.
- Per-route override: chat can be cloud while summaries stay local, or any other combination — see the AI provider reference.
What this looks like for you
24 GB+ Mac
100% local
Everything on-device, including the language model. Default. Nothing leaves your Mac, ever.
16 GB Mac
Hybrid (recommended)
Local recording & transcription, cloud summaries & chat through OpenRouter with zero-retention enforcement.
Any Mac
Per-step control
Pick local or cloud for each pipeline step independently. Set a monthly budget; auto-fall back to local when hit.
Where this is going
Local models are getting better fast — smaller, faster, and smarter on the same hardware. The bar for “good enough on a base Mac” keeps moving down. We expect 100% on-device to be the comfortable default for everyone before long, and we'll keep upgrading the local pipeline as new model families ship. Until then, hybrid is the realistic option that lets you use Hapi well today without compromising on the things that actually matter — your audio and your data — staying on your Mac.
Questions? Email hello@speakhapi.com — we read every one. For the technical reference, see How it works → AI provider. For data-handling specifics, see the privacy policy.