Guide

Which Whisper model do you actually need?

Tiny, Base, Small, Medium, Large. Bigger is more accurate and much slower, and for a lot of recordings the difference is smaller than the marketing suggests.

25 August 2026

Every local transcription app is running some version of OpenAI's Whisper, and most of them make a point of which model sizes they support. Larger models are genuinely more accurate. They are also dramatically slower and much larger on disk, and the accuracy gap depends far more on your recording than on the model.

Here is how to think about it without buying the biggest thing on offer by reflex.

The sizes #

Whisper comes in five main sizes. The numbers matter less than the shape of the curve:

Model	Parameters	Roughly on disk
Tiny	39 million	~75 MB
Base	74 million	~140 MB
Small	244 million	~470 MB
Medium	769 million	~1.5 GB
Large	1.5 billion	~3 GB

Each step up is roughly three times the size of the one below it, and takes correspondingly longer to run. Going from Small to Large is about six times the parameters — on a long recording that is the difference between making a coffee and going for lunch.

The accuracy curve does not rise nearly that fast. It climbs steeply from Tiny to Small, then flattens.

What actually decides your accuracy #

This is the part that gets lost. The quality of your recording matters more than the model size.

A clean recording — one speaker, decent microphone, quiet room, standard accent — will transcribe well on Small. Pushing that same file through Large gets you a handful of corrected words, mostly proper nouns.

The things that genuinely break transcription are:

• Background noise. A café, an air conditioner, traffic.
• Overlapping speech. Two people talking at once is hard for every model.
• Distance from the microphone. Audio recorded across a meeting room table.
• Strong or less-common accents, especially in a language the model saw less of during training.
• Technical vocabulary, names, and acronyms. No model size fixes an unusual surname.
• Languages other than English. Whisper's training data is heavily English-weighted, so the gap between model sizes is much wider for other languages.

If your audio has several of those, a larger model is worth real money. If it has none of them, you are paying for a rounding error.

A practical way to choose #

Tiny — Rough drafts and searchability. Good enough to find which part of a two-hour recording you need. Not good enough to publish.

Base — Voice memos, notes to yourself, clean single-speaker audio where you will read the result and know what you meant.

Small — The sensible default for most work. Interviews on a decent microphone, podcast episodes, lecture recordings, meeting audio where people are close to the mic. This is where the curve flattens for clean English speech.

Medium and Large — Difficult audio, non-English languages, or anything you will publish without reading in full. Also the right choice if your time is worth more than the wait, and re-reading a transcript costs you more than the extra processing.

Just test it #

The honest answer is that nobody can tell you from a description of your recordings, because the variables above interact.

Take your worst file — the noisy one, the one with the difficult accent, the one you already know is a problem — and run it through whatever model sizes you have available. Compare the transcripts. That single test will tell you more than any benchmark table, because it uses your microphone, your room and your subject matter.

If the small model handles your hardest file, it will handle everything else, and you have just saved yourself several gigabytes and a lot of waiting.

Where this app sits #

Being straightforward about it: Offline Transcription ships Tiny, Base and Small. It does not include Medium or Large.

That covers the majority of everyday transcription — clean interviews, podcasts, lectures, meetings, voice memos — and it keeps the app small and fast, including on iPhone. If you routinely work with difficult audio or non-English recordings where the large models genuinely earn their keep, an app that offers them is a better fit, and our MacWhisper comparison says so directly.

The app is free to download, so the test above costs you nothing but the time to run it.

One thing that is not about model size at all #

Whichever model you use, where it runs is a separate question. Whisper is an open model, and it can be run on your own machine or on somebody's server. A cloud transcription service is probably running Large — and also holding a copy of your recording.

Running Small locally and running Large in someone else's data centre are not points on the same scale. If the recording is confidential, that decision comes first, and the model size comes second. Our guide on checking whether an app is really offline covers how to verify it either way.