WhisPaste ships three quantised Whisper models plus one processor-first alternative. Tell the picker what graphics memory you have and it applies the same rule the app itself uses to recommend a tier.
Short answer: start with balanced if you have about 2 GB of graphics memory or more, compact if you have less or no GPU at all, premium only if you have roughly 4 GB of dedicated graphics memory to spare. All three run fully offline, all three cover 99 languages, and you can switch tiers at any time — the app re-downloads a single file.
WhisPaste deliberately offers three tiers rather than the full whisper.cpp catalogue. Each is a quantised GGML build, which is why the downloads stay well under a gigabyte while keeping the quality of the underlying model. You pick a tier in the settings; the app downloads one file, verifies its checksum and switches over.
| Tier | Model | Download | Graphics memory |
|---|---|---|---|
| Compact | whisper-small (q5_1) | 181 MB | ~900 MB |
| Balanced | whisper-medium (q5_0) | 514 MB | ~1.5 GB |
| Premium | whisper-large-v3-turbo (q5_0) | 547 MB | ~2.6 GB |
| Parakeet | parakeet-tdt-0.6b-v3 (int8) | 639 MB | processor only |
JavaScript is off, so all tiers are described below. The comparison table above gives you the same information.
A 181 MB download that needs about 900 MB of graphics memory, and the tier that keeps a modest machine responsive. It is the right starting point on integrated graphics, on an older card, and on anything without a GPU at all. Accuracy is noticeably below the larger tiers on difficult audio, but for clear speech into a decent microphone the gap is smaller than the download size suggests.
A 514 MB download that wants about 1.5 GB of graphics memory. This is the tier most people should be on: clearly better than compact on accents, technical vocabulary and background noise, without the memory appetite of the top tier. On Apple Silicon an 8 GB machine already carries it.
A 547 MB download that wants about 2.6 GB of graphics memory. The turbo variant of large-v3 is the most accurate model WhisPaste ships and, thanks to its reduced decoder, far quicker than a plain large model. Worth it if you have a dedicated card with roughly 4 GB or more, and it is the tier to pick when your audio is hard: fast speech, mixed languages, unfamiliar names.
Parakeet is the second on-device engine, not a Whisper tier. It is a 639 MB download, has no GPU backend at all, and is still several times faster than Whisper on the same processor — which makes it the strongest option on a machine without usable graphics acceleration. The trade-off is language coverage: 25 European languages instead of Whisper’s 99.
The picker above is not a guess — it applies the same thresholds the app uses. Dedicated NVIDIA cards are judged against roughly 70 % of their reported memory, because the operating system and other processes take their share: about 2.1 GB reported to reach balanced, about 3.7 GB to reach premium. Apple Silicon is treated generously, since unified memory is genuinely shared: 2 GB for balanced, 4 GB for premium. Integrated AMD and Intel graphics are treated the most cautiously of all — they report borrowed system memory as if it were their own, so premium only gets recommended from about 12 GB upward.
Before optimising the tier, check the language. Whisper covers 99 languages at every tier, so a larger model does not unlock a new language — it only transcribes the same language more accurately. Parakeet is the opposite trade: much faster on a processor, restricted to 25 European languages including German, English, French, Spanish, Italian, Polish, Dutch, Ukrainian and Russian. If your language is on that list and your machine has no usable GPU, Parakeet will beat any Whisper tier on speed.