Fundamentals

Whisper Model Sizes Explained: Which One Should You Use?

OpenAI's Whisper comes in a family of sizes, from tiny to large-v3, plus the faster large-v3-turbo. Bigger models get names, jargon and accents right more often; smaller ones are faster and need less memory. Here is what each size is, how they compare, and how to pick the right one for your own computer.

We build AirTypes, which runs the two smallest of these models for you. Its Plain mode is free and offline: Whisper runs on your CPU and nothing is uploaded. If you want more accuracy than a small model gives, or your laptop struggles, Polished skips local models entirely and transcribes and cleans up your words in the cloud, 120 minutes for every dollar. Here is what Plain typed, offline, from a real sentence:

You say

"ok so um the meeting is moving to tomorrow at ten, client is not available, and uh send him the invoice again"

→
Plain types (Whisper on your CPU)

The meeting is moving to tomorrow at ten, client is not available, and uh send him the invoice again.

The Whisper model sizes are tiny (about 39 million parameters), base (74M), small (244M), medium (769M), large-v3 (about 1.55 billion) and large-v3-turbo (809M). They all do the same job, turning speech into text, but larger models have more capacity to learn, so they make fewer mistakes on unusual words and difficult audio. The cost is speed and RAM: each step up takes longer to transcribe and needs more memory.

If you run Whisper yourself, small or medium is the sweet spot for most people dictating in English on a recent laptop. Pick large-v3-turbo if you have 8 GB of RAM and a 6-core CPU and want near-top accuracy, and large-v3 only if accuracy matters more than anything else and your machine has room for it.

Short answer

  • Old or low-memory laptop: tiny.en or base.en.
  • Everyday writing on a normal laptop: small.en.
  • Technical vocabulary or long dictation: medium.en or large-v3-turbo.
  • Best accuracy, hardware no object: large-v3.
  • Laptop too slow for any of these: skip local models and use a cloud option such as AirTypes Polished.
  • In AirTypes: Plain ships tiny.en and base.en; for more accuracy it uses Polished on the server, not a bigger local model.

What the Whisper model sizes are

Whisper is an open speech recognition model that OpenAI released in 2022. It is an encoder-decoder transformer: the encoder listens to the audio (turned into a spectrogram), and the decoder writes out the text one token at a time, with punctuation and capital letters included. OpenAI trained it at several sizes so it can run on anything from a small laptop to a workstation. If you want the bigger picture of how this works without the cloud, our offline speech recognition explainer covers the whole pipeline.

The main sizes

The sizes differ in how many layers the network has and how wide each layer is, which sets the parameter count. More parameters means more learned detail, and also more maths to run for every second of audio.

  • tiny (~39M parameters): the smallest. Very fast, but it trips over uncommon words.
  • base (~74M): still light, noticeably steadier than tiny.
  • small (~244M): the first size that feels dependable for ordinary writing.
  • medium (~769M): handles longer sentences and specialist terms well.
  • large-v3 (~1550M): the most accurate release, and the heaviest.
  • large-v3-turbo (~809M): a faster version of large-v3, explained below.

What makes large-v3-turbo different

Turbo is not a separate model trained from scratch. It is large-v3 with a much smaller decoder: 4 decoder layers instead of 32. The encoder, which does most of the listening, stays the same size. Because the decoder runs once for every token it writes, shrinking it makes transcription much faster. OpenAI describes the accuracy cost as minimal, and in practice it lands close to large-v3 for everyday speech.

English-only .en variants

tiny, base, small and medium also come as English-only models: tiny.en, base.en, small.en and medium.en. They are the same size but focus all their capacity on English. OpenAI's notes say the .en versions tend to do better on English, with the biggest difference at tiny and base, shrinking at small and medium. There is no .en version of the large models.

The trade-off: accuracy vs speed and RAM

Every size can transcribe a clear, simple sentence like "send the report on Friday." The differences show up at the edges:

  • Names and jargon. Smaller models fall back on common words when they are unsure, so "Kubernetes" or a colleague's surname may come out as something that merely sounds similar. Larger models have seen more and guess better.
  • Accents and fast speech. Larger models are more forgiving of accents, mumbling and speed changes.
  • Long sentences. Smaller models are more likely to lose punctuation or drop a word partway through a long, winding sentence.
  • Noise. Background sound affects every model, but the small ones degrade first.

On the other side of the scale, every step up costs you. A larger model takes longer to download, uses more RAM while it runs and takes longer to return text after you stop speaking. On a machine that is short on memory, a model that is too big can slow everything else down too. The best model is not the biggest one; it is the biggest one your computer runs comfortably.

Whisper model comparison table

Parameter counts are OpenAI's published figures. Download sizes are for the whisper.cpp model files that desktop apps run on the CPU (AirTypes Plain uses the first two); hardware suggestions are for CPU-only use.

ModelParametersDownloadGood forSuggested hardware
tiny.en~39M~78 MBQuick notes, short messages, very old machinesAny device
base.en~74M~148 MBEveryday chat and email on light hardwareAny device
small.en~244M~488 MBMost daily writing, good balance4 GB RAM
medium.en~769M~1.5 GBLonger documents, some technical terms6 GB RAM, quad-core
large-v3-turbo~809M~1.6 GBNear-top accuracy with a fast turnaround8 GB RAM, 6-core
large-v3~1550M~3.1 GBDense jargon, accents, highest accuracy12 GB RAM, 8-core

Which Whisper model should you use?

Start from your situation, not from the biggest number. This list is for running Whisper yourself, with any tool; AirTypes Plain ships two of these sizes. Find the line that fits you best:

  1. 1
    Old laptop or under 4 GB of RAM

    Use tiny.en or base.en. They stay quick on almost anything. Expect to fix the odd name, and speak a little more clearly than usual.

  2. 2
    Everyday writing: email, chat, notes

    Use small.en. On a machine with 4 GB of RAM or more it is the best balance of speed and accuracy for ordinary language.

  3. 3
    Technical vocabulary: code, medicine, law, research

    Use medium.en, or large-v3-turbo if you have 8 GB of RAM and a 6-core CPU. Both recognise specialist words far more often than the small sizes.

  4. 4
    Long-form dictation: articles, reports, chapters

    Use large-v3-turbo. It keeps long, complex sentences intact and still returns text quickly, which matters when you dictate for an hour.

  5. 5
    Best possible accuracy

    Use large-v3 if you have 12 GB of RAM and an 8-core CPU. It is the slowest option, so save it for work where every word counts.

If you are unsure, start one size below what you think you need. If it mishears words you use often, step up. If you wait too long for text to appear, step down.

Try Whisper without the setup

AirTypes Plain downloads tiny.en or base.en for you and runs it offline on your CPU. Free on macOS and Linux, no card, no time limit.

Download AirTypes

How AirTypes maps the sizes to tiers

The AirTypes desktop app runs on macOS and Linux. You hold a hotkey (Ctrl + Shift + Space by default), speak, let go, and the text is typed at your cursor in any desktop app. In Plain mode, Whisper runs on your own computer. AirTypes Plain ships two of the sizes, and instead of asking you to remember model names, it gives each a tier name:

  • Swift: tiny.en, ~78 MB. The fastest.
  • Crisp: base.en, ~148 MB. The default the app suggests, and the tier we recommend.

Plain is English only. Swift and Crisp use the English-only .en variants, which focus their capacity on English. Both run offline on your CPU with no GPU, on a 4 GB machine, and 0 bytes of audio are uploaded. You can switch tiers any time: each model downloads once and then stays on your device. For more accuracy than Crisp, on technical terms, accents, long documents or any language other than English, AirTypes uses Polished on the server instead of a bigger local model. The AirTypes setup guide walks through choosing a tier on first launch.

Plain also does light cleanup on the machine: it removes a leading um or uh, capitalizes the first letter and adds a period. It does not rewrite your sentence, so a filler in the middle stays, as in the example at the top. Platform guides cover the details for dictation on macOS and Whisper dictation on Linux. The Windows desktop app is in development; on Windows today you can Dictate in your browser (new in November 2026), which is Polished only.

Or skip local models entirely: Polished

If your laptop is old, short on RAM, or already busy with other work, or you want more accuracy than a small model gives, picking a model size may not be the right question. Even Swift has to run on your CPU. The alternative is to let a server do the heavy lifting.

That is AirTypes Polished. Nothing heavy runs locally: Polished sends the recording to AirTypes to transcribe and clean it up. A large speech model transcribes it on our servers, then an AI removes filler words and false starts and fixes grammar and punctuation, keeping your meaning, numbers and names. A guard keeps your numbers and catches rewrites that drift too far from what you said; when it trips, you get your own words instead. With Always translate to English on, the guard checks numbers only, so check names. A real example:

You say

"so um I think we should uh probably move the deploy to friday because like the migration isn't, the migration isn't tested yet and uh QA needs two more days"

→
Polished types

I think we should move the deploy to Friday because the migration isn't tested yet and QA needs two more days.

Polished also works in the language you speak, not just English (English, Hindi, Bangla, Japanese, German and most other spoken languages; try yours). The trade-offs: it needs an internet connection, and the recording leaves your machine. It is prepaid: $5 buys 600 minutes of speech, credit never expires, and there is no subscription. When credit runs out, Polished pauses and Plain keeps working. You switch modes on the Home screen. See how Polished works.

Tips that help accuracy on any model

A bigger model is not the only lever. These habits help tiny and large alike:

  • Use a decent microphone. A headset or a USB mic close to your mouth beats a laptop's built-in mic across the room. Clean audio does more for accuracy than one step up in model size.
  • Speak in full sentences. Whisper uses context to choose between similar-sounding words. A complete thought gives it that context; three isolated words do not.
  • Pause briefly before you start and after you finish. A short gap after pressing the hotkey, and before releasing it, keeps the first and last words from being clipped.
  • Find a quiet room. Fans, music and other voices compete with yours. Turn off what you can.
  • Keep a steady pace. You do not need to speak slowly, just evenly. Rushing through a long list is where words get dropped.

If you are new to dictation, building the habit matters as much as the model. Our guide to your first 7 days of voice dictation covers that side.

FAQ

What is the most accurate Whisper model?

Large V3 is the most accurate Whisper model OpenAI has released for transcription. It is also the largest and slowest, with about 1.55 billion parameters. Large V3 Turbo comes close at a much higher speed, which is why it is the better choice for many people.

Is Whisper Large V3 Turbo worse than Large V3?

Slightly. Turbo keeps the Large V3 encoder but cuts the decoder from 32 layers to 4, so it transcribes much faster. OpenAI describes the accuracy loss as minimal. For everyday dictation most people will not notice a difference; for dense jargon, heavy accents or long passages, full Large V3 can still catch a few more words.

Do I need a GPU to run Whisper?

No. A GPU speeds up the larger models, but Whisper runs on an ordinary CPU; larger models simply need more RAM and more CPU cores to feel quick. AirTypes Plain ships two of the sizes, Swift (tiny.en) and Crisp (base.en), and runs them on your CPU with no GPU required. For more accuracy than Crisp, AirTypes uses Polished on the server instead of a bigger local model.

What is the difference between tiny and tiny.en?

tiny is multilingual: it can transcribe a wide range of languages and translate into English. tiny.en is the same size but English only. According to OpenAI, the .en variants tend to do better on English, and the gap is largest at the tiny and base sizes. English-only versions exist for tiny, base, small and medium, not for the large models.

Get AirTypes, free

Plain runs the Whisper tier you pick offline on your CPU, free with no time limit. Polished is optional when you want cleaner text in your language. See pricing. macOS and Linux now. The Windows desktop app is in development; on Windows, Dictate in your browser today.

Download for macOS and Linux