Source: https://saypad.app/which-dictation-model-mac

# Which dictation model should you use on your Mac?
Saypad's model library lists 18 Whisper models, from a 32 MB file to a 3.1 GB one. Most people never need to open it: the app picks a good default for you. If you're short on disk space, dictate only in English or just want to understand the names, this guide explains what each one means and what our tests show.

## The short answer

Leave it on the default. Saypad ships with a small model inside the 187 MB installer, so you can dictate right away, even offline. The more accurate large-v3-turbo model in its 8-bit version (874 MB) then downloads in the background on Wi-Fi, and Saypad switches to it by itself. That's the model we recommend for nearly everyone. The rest of the library is there for people with a specific reason to change it.

## How to read the names

**Size: tiny, base, small, medium, large.** Bigger models know more about speech and get more words right. They also take more disk space and more time to recognize each dictation.

**large-v3, large-v2 and turbo.** large-v3 is the newest full-size model, and large-v2 the one before it. OpenAI describes turbo as an optimized version of large-v3, and large-v3-turbo is the one Saypad downloads by default. OpenAI's own speed table, measured on a data-center GPU rather than a Mac, puts turbo at about eight times the speed of large.

**.en** means English only. OpenAI says these tend to do better on English, especially tiny.en and base.en. They can't recognize other languages, so skip them if you ever dictate in anything else. The models without .en detect the language of each dictation on their own.

**q5 and q8** mean compressed. The numbers inside the model are stored with fewer bits (5 or 8), which makes the file smaller. Sometimes that costs accuracy and sometimes it doesn't. Our test below shows both cases.

## All 18 models

|  | Size | Library note |
|---|---|---|
| large-v3-turbo, 8-bit (q8_0) | 874 MB | Recommended: accurate and fast |
| large-v3-turbo | 1,625 MB | Turbo, full precision |
| large-v3-turbo-q5_0 | 574 MB | Turbo, smaller file |
| large-v3 | 3.1 GB | Most accurate, slowest |
| large-v3-q5_0 | 1,081 MB | Large v3, smaller file |
| large-v2 | 3.1 GB | Previous large model |
| medium | 1,534 MB | Balanced |
| medium.en | 1,534 MB | Balanced, English only |
| medium-q5_0 | 539 MB | Medium, smaller file |
| small | 488 MB | Fast, good for clear speech |
| small.en | 488 MB | Fast, English only |
| small-q5_1 | 190 MB | Built into the app |
| base | 148 MB | Very fast, rough |
| base.en | 148 MB | Very fast, English only |
| base-q5_1 | 60 MB | Base, smallest file |
| tiny | 78 MB | Fastest, for quick notes |
| tiny.en | 78 MB | Fastest, English only |
| tiny-q5_1 | 32 MB | Tiny, 32 MB |

Sizes are what each file takes on disk. The notes are the descriptions shown in the library.

## What we measured

We ran 32 real dictations through four of these models on a MacBook Pro with an M4 Pro chip and compared the text word by word against the full large-v3-turbo. Most recordings were Russian with English terms mixed in.

- **large-v3-turbo, 8-bit (874 MB):** the same text as the full model in 31 of 32 dictations, and the one difference was punctuation. Median time per dictation 1.98 s from a cold start, against 2.14 s for the full model.
- **large-v3-turbo-q5_0 (574 MB):** identical in 21 of 32, with 2.4% of words different, and no faster (2.21 s).
- **small-q5_1 (190 MB, built in):** about 2.5 times faster (0.84 s), but 22% of words came out different and no dictation matched exactly.

In an earlier test, tiny and base turned Russian into nonsense. We haven't benchmarked medium, large-v3, large-v2 or the .en models, so we won't put numbers on them. In daily use the model stays loaded, so a 7.6-second sentence appears about 1.2 seconds after you let go of the key with the default model. The full write-up is in [how accurate offline dictation is on a Mac](offline-dictation-accuracy-mac).

## When to pick something else

- **Short on disk space:** stay on the built-in small model and turn off the background download. Fine for quick messages, but expect to fix words, especially names.
- **Only dictate in English and want a small file:** try small.en or base.en instead of their multilingual versions. We haven't tested them against turbo, so compare on your own voice.
- **Curious about the largest model:** large-v3 is 3.1 GB, and the library describes it as the most accurate and slowest. Try it and switch back if dictation feels slow.
- **Mix languages:** use a model without .en. [Dictating in two languages](dictate-multiple-languages-mac) explains how detection works.

## Switching, downloading and deleting

Open Settings → Recognition → Model library. Download any model, switch with one click and delete the ones you don't need. The built-in small model can't be deleted, so dictation always has something to run. A paused download resumes only for the same model, so switching mid-download can't mix files. Every model runs on your Apple Silicon Mac through whisper.cpp with Apple's Metal graphics, and nothing you say leaves the Mac. On macOS 26 there's also Apple's on-device speech engine as a second option. For how this compares with other local tools, see [offline speech to text on a Mac](offline-speech-to-text-mac).

## Questions

**Which model should I use for dictation on a Mac?**
For most people, the default: large-v3-turbo in its 8-bit version (874 MB). In our test on 32 real dictations it gave the same text as the full 1.6 GB model in 31 of them and was slightly faster.

**Are the English-only models better for English?**
OpenAI says the .en models tend to perform better for English-only use, especially tiny.en and base.en. We haven't measured them in Saypad, and they can't recognize any other language, so try one on your own voice before switching.

**Can I delete models I don't use?**
Yes. In Settings → Recognition → Model library you can delete any downloaded model. The small model built into the app stays, so dictation always works.

**Do bigger models need the internet?**
Only to download them. Every model in the library runs on your Mac, and each one is checked before use. After that, dictation works offline.