Which Whisper model to use on an Apple Silicon Mac
whisper.cpp offers a long list of model files, and the names don't explain much: tiny, base, small, medium, large-v3, large-v3-turbo, with .en and q5_0, q5_1 or q8_0 tacked on. FnFlow's model library lists 18 of them, from tiny at 32 MB to large-v3 at 3.1 GB. Here's how to choose a Whisper model for dictation on a Mac, using only numbers we measured ourselves.
Reading the names
- Size: tiny, base, small, medium and large. Bigger is more accurate and slower.
- large-v3-turbo is a faster variant of large-v3 and the usual pick for dictation.
- .en means English-only. These exist for tiny, base, small and medium. There's no .en version of large-v3-turbo.
- q5_0, q5_1, q8_0 are quantized files: weights stored in 5 or 8 bits instead of 16, so the download is smaller. The question is how much accuracy each one costs.
What we measured
On 8 October 2026 we ran 32 real dictations, 339 seconds of audio, through four models on a MacBook Pro with M4 Pro, using whisper.cpp 1.9.4 with Metal. The audio was mostly Russian with English technical terms, from one speaker. We compared each model's words with the full f16 model's output, so these numbers show what you lose against the best model, not absolute accuracy. Every timing is a cold run that includes loading the model.
- large-v3-turbo f16, 1,625 MB: the reference. Median 2.14 s, 90th percentile 2.33 s.
- large-v3-turbo q8_0, 874 MB: 0.0% word difference, same text in 31 of 32 files. Median 1.98 s, p90 2.09 s.
- large-v3-turbo q5_0, 574 MB: 2.4% word difference, same text in 21 of 32. Median 2.21 s, p90 2.86 s.
- small q5_1, 190 MB: 22.0% word difference, never identical. Median 0.84 s, p90 1.02 s.
The one q8_0 file that differed only changed punctuation and case. The full write-up and CSV are in our whisper.cpp benchmark on Apple Silicon.
q8 is free, q5 isn't
q8_0 halved the download and gave the same words, and it was a little faster at both median and p90. q5_0 saved another 300 MB, but it changed at least one word in 11 files and the slow tail got worse. In a dictation app, where you read every sentence before sending it, a wrong word costs more than 300 MB of disk. We don't know for sure why q5_0 ran slower. Our guess is that dequantizing 5-bit weights on Metal costs more than the smaller read saves.
The small end
An earlier run on 32 recordings, 24 Russian and 8 English, put the small models against large-v3-turbo. tiny differed by 52% of words on Russian and 65% on English. base was at 35% and 65%. Both produced nonsense on Russian. small q5_1 was at 24% and 41%, and it's the smallest model we'd ship. The English sample was small and mixed with Russian, so treat those English figures loosely.
Small models still have a job. small q5_1 runs at under a tenth of real time, so it's fine as a first model while a larger one downloads, or for quick notes where you'll fix a word or two.
Cold vs warm
Load time matters as much as model size. Kept loaded in a local whisper-server, large-v3-turbo returned a 7.6-second sentence in 1.17 to 1.19 seconds, against 2.26 seconds for a cold run on the same audio. If you script Whisper yourself, keep the model resident. The setup is in how to run Whisper locally on a Mac.
Our picks
- Default: large-v3-turbo q8_0. Same text as f16 in our test, 874 MB.
- Tight on disk or just starting: small q5_1, then switch up when you can.
- English only: try a .en model against q8_0 on your own voice. We haven't measured them.
- Maximum accuracy, plenty of disk: large-v3 at 3.1 GB is in the library, but we haven't benchmarked it, so check the speed on your Mac first.
Your numbers will move with the chip, the language and how clearly you speak. If you see phantom phrases on silence, that's a separate problem, covered in why Whisper says "Thank you".
Switching models in FnFlow
FnFlow ships small q5_1 inside the installer, so dictation works offline right after install. large-v3-turbo q8_0 downloads in the background on Wi-Fi, and the app switches to it when it's ready. Settings → Recognition lists all 18 models. You can download any of them, switch with one click and delete the ones you don't use, except the built-in small model. Every model runs on your Mac and is checked before use. What FnFlow doesn't do is pick for you per dictation. One model is active at a time.
Looking beyond Whisper? See Parakeet vs Whisper on a Mac and Apple SpeechAnalyzer vs Whisper.
Questions
What's the best Whisper model for dictation on a Mac?
In our test, large-v3-turbo q8_0. It gave the same text as the full f16 model in 31 of 32 dictations at about half the download (874 MB) and was slightly faster.
Should I use an English-only .en model?
If you only speak English, they're worth a try. .en versions exist for tiny, base, small and medium, but not for large-v3-turbo. We haven't benchmarked the .en models, so we can't give you numbers.
Is q5 quantization good enough?
Not in our test. large-v3-turbo q5_0 saved 300 MB over q8_0 but changed words in 11 of 32 files and had a slower 90th percentile (2.86 s vs 2.09 s).
Why is the first dictation slow?
On the first run of a new build, Metal compiles its pipelines. That took 15 to 17 seconds on our M4 Pro. FnFlow does it in the background after launch.
Local Wispr Flow alternativeFnFlow vs VoiceInkVoice input for CursorHow 7 Mac dictation apps get you to your first dictationWhisper.cpp on Apple Silicon: which large-v3-turbo quantization to shipHow I built a local dictation appVoice dictation in Claude Code and the terminalWhy holding fn doesn't work for dictation on Mac (and how to fix it)How to run Whisper locally on an Apple Silicon MacSpokenly alternative for developersFnFlow vs EnviousWisprHow to dictate a good prompt for AI coding toolsVoice dictation in VS Code on a MacTeaching Whisper your jargon: initial prompt and replacementsWhy Whisper types "Thank you" on silence, and how to filter itPrivate dictation on a Mac at work: what actually leaves your machineVoice input for Codex CLI and the ChatGPT appVoice coding on a Mac: what actually worksWriting git commit messages by voiceDictating code review comments on GitHubDictation in IntelliJ IDEA, PyCharm and WebStorm on a MacDictation in Zed on a MacDictation in Terminal, iTerm2 and WarpPush-to-talk dictation on a Mac: hold, toggle or streamSuperwhisper alternative for developersClaude Code voice mode: how to enable /voice and fix itWhisper apps: five ways to run Whisper locallyTalon voice coding vs a dictation appPush to talk on a MacDoes Claude have a voice mode?VoiceInk vs Wispr FlowVoiceInk vs SuperwhisperParakeet vs Whisper on a MacApple SpeechAnalyzer vs WhisperWhisper.cpp vs faster-whisper on a MacMlx-whisper vs whisper.cppGitHub Copilot voice inputGemini CLI voice inputVoice input in Google AntigravityVibe coding with voice input