Why we built our own model.
Apple Foundation Models, the on-device LLM Apple ships in macOS 26, is already remarkable: fast, private, and free of the cloud round-trip that defines every other AI tool. But it's a generalist. It doesn't know what a clean dictation looks like, when to drop your "ums," how to keep a Linear ticket terse and a Mail draft warm, or that "ping the team in Slack" should always feel a little excited.
Most apps would solve that with a cloud LLM. We took the harder path: train a small adapter that runs on top of Apple FM, on the same Neural Engine, with no third party in the loop.
How it works.
Dollop AI is a LoRA-style adapter, a thin layer of weights specific to our use cases, bolted onto Apple's foundation model at runtime. Apple already keeps the base model warm in memory; our adapter is tiny by comparison and loads in milliseconds when you press ⌥ Space or ⌃ A.
The result feels like Apple FM, but tuned for the way people actually speak into a Mac: messy, mid-thought, with names and product vocabulary the generalist doesn't know.
Faster, sharper, and the RAM stays at zero.
Because the adapter sits on top of the model Apple already loads, Dollop adds nothing to your RAM budget. Compare that to local Whisper-based tools (700 MB–2 GB resident) or cloud tools (your audio leaves your Mac entirely):
- Dollop AI: 0 GB added · on-device · Apple Neural Engine
- Local Whisper apps: 0.7–2 GB resident · on-device
- Cloud dictation (Wispr Flow, Cluely): audio routed off-device
Latency lands in the same range as raw Apple FM. First token in tens of milliseconds, full response in well under a second for typical dictations. The adapter is small enough not to slow the base model down.
What it actually changes.
- Cleanup that knows context: "ping" stays casual in Slack, formal in Mail, terse in Linear.
- Better names and jargon: "ARR," "APAC," "Sora," "Anthropic" are recognized verbatim instead of phonetically.
- Tighter AI Mode answers: shorter, more direct, less LLM–preamble.
- Chat overlay that sticks to your topic: trained on real screen-context conversations, not generic web chats.
Still on-device. Still free.
Dollop AI ships inside the app. No subscription, no usage tier, no account required. The adapter runs entirely on your Mac, on Apple's Neural Engine, with no audio or transcript leaving the machine. The privacy posture is the same as Apple FM itself, because that's exactly what it's running on.