Whisper — OpenAI's open-source speech recognition model — comes in six sizes, from a 39 MB Tiny model to a 1.5 GB Large. Every Mac dictation and transcription app built on it (including Tota) makes you choose one, and the names don't tell you much. Here's what actually changes between sizes, and how to pick.
TL;DR
The Whisper Family at a Glance
| Model | Parameters | Download | Best for |
|---|---|---|---|
| Tiny | 39M | ~39 MB | Oldest/lowest-RAM machines; speed over accuracy |
| Base | 74M | ~74 MB | A solid default; ships built into Tota |
| Small | 244M | ~244 MB | The dictation sweet spot on Apple Silicon |
| Medium | 769M | ~769 MB | Harder audio when you can spare the RAM |
| Large v3 | ~1.55B | ~1.5 GB | Maximum accuracy; multilingual + translation |
| Large v3 Turbo | ~809M | ~1.5 GB | Large-class accuracy, several times faster |
Two structural things worth knowing. First, every size from Tiny to Medium comes in two flavours: multilingual, and an English-only .en variant that spends all of its capacity on one language — at the same size, the .en model transcribes English better, and the gap is biggest at the small end. Second, Large v3 Turbo is not just "Large but newer": OpenAI pruned the decoder from 32 layers to 4 and fine-tuned the result, so it decodes several times faster than Large v3 while giving up only a little accuracy. For most people it makes full Large v3 obsolete outside of translation and edge-case audio.
What "Bigger" Actually Buys You
Accuracy improves with size, but not linearly — the curve flattens hard after Small. On clean audio (one speaker, decent microphone, no background noise), Small already gets most sentences exactly right, and a Large model mostly corrects errors you'd have caught at a glance. Where the big models genuinely earn their memory:
- Accents and dialects far from the broadcast-English centre of the training data
- Jargon and proper nouns — drug names, case law, API identifiers
- Bad audio — echoey rooms, phone recordings, crosstalk, background music
- Non-English speech and translation, which only the Large models handle well
If none of those describe your audio, a smaller model plus a domain-aware correction layer beats a bigger model alone — that's the architecture argument made on our AI dictation for Mac page, where a compact local language model fixes exactly the vocabulary a mid-size Whisper model misses.
What Your Mac Can Handle
On Apple Silicon, model weights live in unified memory alongside everything else you're running, so the real question is RAM headroom, not chip generation:
- 8 GB Macs: Tiny, Base, and Small all run comfortably. Small is the sensible ceiling for an always-loaded dictation model; a 1.5 GB Large download will run, but you'll feel it when Xcode or Chrome wants the memory back.
- 16 GB Macs: Medium is comfortable full-time, and Large v3 Turbo is fine for file transcription sessions.
- 32 GB and up: Run whatever you like; Turbo full-time is realistic.
Latency is the other axis. Dictation lives and dies on the pause between releasing the record key and seeing your words — a short utterance through Small feels instant on an M-series Mac, while full Large v3 adds a beat of thinking time. File transcription doesn't care about that first-response feel, only total throughput, which is exactly why Turbo exists.
How Tota Uses These Models
Tota runs Whisper through WhisperKit, Argmax's open-source Core ML framework, which compiles the models for Apple's Neural Engine and GPU. Base (English) ships inside the app so dictation works before you've downloaded anything; the other five tiers are a one-time download from Settings, offline thereafter. The same loaded model powers both live dictation and Audio Drop file transcription, and the multilingual Large models unlock 13 languages plus translate-to-English — details in the models & languages documentation.
Honest note: if you want to experiment with raw Whisper — custom prompts, fine-tunes, batch scripting — a dedicated tool like whisper.cpp or MacWhisper is the better laboratory. Tota's job is making a well-chosen model disappear into your daily typing.
The Bottom Line
Start with what's already there (Base), move to Small (English) the first time a technical term comes out mangled, and keep Large v3 Turbo on hand for transcribing recordings. Everything runs on your Mac either way: no audio leaves the machine, and the model you downloaded works the same on a plane as on office WiFi.
Frequently asked questions
What's the difference between Whisper Large v3 and Large v3 Turbo?
Large v3 Turbo is a pruned, fine-tuned version of Large v3: OpenAI cut the decoder from 32 layers to 4 and retrained it, shrinking the model from about 1.55 billion parameters to about 809 million. Decoding is several times faster with only a small accuracy cost, which is why Turbo is usually the better pick of the two on a Mac.
Is a bigger Whisper model always more accurate?
Accuracy rises with size, but with steep diminishing returns. For a clear speaker dictating English into a decent microphone, Small gets most sentences exactly right and Large mainly earns its size on hard audio: heavy accents, technical jargon, background noise, and crosstalk. Match the model to your audio, not to the biggest number.
Can an 8 GB Mac run Whisper Large?
Often yes, but it's tight — a ~1.5 GB model plus inference overhead competes with everything else in unified memory. On an 8 GB Apple Silicon Mac, Small is the comfortable ceiling for always-on dictation; load Large v3 Turbo for a one-off file transcription and close it afterwards.
What do the .en (English-only) Whisper models do?
Tiny through Medium come in multilingual and English-only (.en) variants. The English-only versions spend their entire capacity on one language, so at the same size they transcribe English better — the gap is biggest at Tiny and Small and fades by Medium. If you only dictate in English, prefer the .en variant.
Do Whisper models work offline?
Yes — completely. A Whisper model is a file of weights that runs locally; after a one-time download it needs no internet connection at all. In Tota, all transcription runs on-device via WhisperKit, and the app works with WiFi off.
