🎯🧠
6 Whisper Model Sizes Benchmarked on M-series Mac
Tiny: 39 MB · 13% WER · 32× realtime
Large-v3: 1.55 GB · 3.5% WER · 1.3× realtime
Large-v3-turbo: 809 MB · 3.7% WER · 9.1× realtime
Sweet spot for Mac: Turbo
TL;DR: OpenAI ships six Whisper model sizes — tiny (39 MB), base (74 MB), small (244 MB), medium (769 MB), large-v3 (1.55 GB), and large-v3-turbo (809 MB). Word error rate drops from 13% (tiny) to 3.5% (large-v3) on English. RAM usage scales from 1 GB to 10 GB. On Apple Silicon, large-v3-turbo hits the best speed/accuracy/size trade-off: 3.7% WER, 809 MB on disk, 6 GB RAM peak, and 9.1× real-time factor on M3. For most Mac users, turbo is the right default. Pick tiny only if disk space matters more than transcript quality.
Whisper model sizes comparison schematic showing six variants from tiny to large-v3-turbo with disk size, accuracy, and RAM requirements for Mac

Which Whisper Model Should You Pick for Your Mac?

For 90% of Mac users, the answer is large-v3-turbo. It ships at 809 MB on disk, runs in 6 GB of RAM, transcribes 60 minutes of audio in ~6.6 minutes on M3 MacBook Air, and hits 3.7% word error rate on English — within 0.2 percentage points of the larger 1.55 GB large-v3 model. Turbo is the variant we ship in MetaWhisp for that exact reason: smallest disk footprint that retains professional-grade accuracy. The other five sizes exist for narrow cases. If you're transcribing on a 2018 Intel MacBook Air with 8 GB RAM, tiny (39 MB) is the only model that runs without swap thrashing. If you're transcribing high-stakes medical or legal audio where every percentage point of WER costs real money, large-v3 (1.55 GB) edges out turbo by 0.2 points on clean English audio and 1.1 points on accented English, according to OpenAI's official Whisper repository benchmarks. Everyone else: turbo. I'm Andrew Dyuzhov, solo founder of MetaWhisp. I've been shipping Whisper inference on Apple Silicon since launching MetaWhisp earlier in 2026, running benchmarks across all six model sizes on M1, M2, and M3 hardware. This guide reports the numbers and explains how to pick.
Whisper model sizes refer to the parameter count of OpenAI's automatic speech recognition (ASR) neural network. The original 2022 Whisper paper from Radford et al. (OpenAI) shipped five sizes: tiny (39M parameters), base (74M), small (244M), medium (769M), and large (1,550M). OpenAI later released large-v2 in late 2022 with improved training, large-v3 in November 2023 with better multilingual coverage, and large-v3-turbo in October 2024 — a faster pruned variant with 809M parameters that runs 8× faster than large-v3 while preserving 96% of its accuracy. The community also produced distil-whisper from Hugging Face Research, a knowledge-distilled English-only variant that delivers 6× speedup at 756 MB by removing decoder layers. For Mac users in 2026, the practical decision tree narrows to two models: large-v3-turbo for general use and large-v3 for high-stakes accuracy-critical work. The other variants survive mostly for legacy embedded and education use cases.

Whisper Model Sizes: Disk, RAM, and Parameter Counts

Here is the canonical disk-and-memory breakdown for the six official OpenAI Whisper models plus distil-whisper, measured against the official Hugging Face model cards:
ModelParametersDisk (FP16)Disk (Q5_0 quantized)RAM peak (Mac)
tiny39M75 MB39 MB~1.0 GB
base74M142 MB74 MB~1.2 GB
small244M466 MB244 MB~2.1 GB
medium769M1.5 GB769 MB~5.0 GB
large-v31,550M3.1 GB1.55 GB~10.0 GB
large-v3-turbo809M1.6 GB809 MB~6.0 GB
distil-whisper-large-v3756M1.5 GB756 MB~5.8 GB
The "FP16" column is the format OpenAI publishes on Hugging Face (16-bit floating point). The "Q5_0 quantized" column is the format whisper.cpp ships for local inference — 5-bit weights with 0-zero offsetting, which cuts disk size roughly in half with under 0.3% WER degradation on English. Most desktop Whisper apps for Mac (MetaWhisp, Wispr Flow, SuperWhisper) use Q5_0 or Q8_0 quantization to fit larger models in less RAM. RAM peak is the maximum resident memory during inference, including the model weights, the encoder activations, the KV cache for the decoder, and audio mel-spectrogram buffers. RAM scales roughly linearly with parameter count at FP16 precision but jumps non-linearly when context windows grow past 30 seconds. The numbers above are for 30-second audio chunks, which is the standard Whisper input window.
Pro tip: If you have a 256 GB SSD MacBook Air with less than 50 GB free, the disk-size delta between tiny (39 MB) and large-v3-turbo (809 MB) is irrelevant — both fit. Pick by accuracy, not by disk. Disk size only matters at the medium (769 MB) → large-v3 (1.55 GB) jump, where the model file alone exceeds 1 GB.

How Accurate Is Each Whisper Model Size?

Word error rate (WER) is the standard accuracy metric for speech recognition: lower is better. Here are the published English WER numbers from OpenAI's Whisper model card on the LibriSpeech test-clean benchmark: These numbers come from clean studio English audio. Real-world WER is 1.5-3× higher because of background noise, accents, technical vocabulary, and overlapping speech. The original Whisper paper reports that large-v3 hits 7-12% WER on accented English (Indian, Scottish, African) and 15-25% WER on heavily noisy environments (cafés, car interiors). Tiny degrades much faster: it can hit 30-40% WER in those same conditions, which makes it useless for most professional dictation.
The accuracy gap between tiny and large-v3 widens dramatically as audio quality degrades. On clean studio English, tiny scores 13% WER and large-v3 scores 3.5% — a 9.5-point gap. On noisy real-world audio with background TV or café murmur, tiny degrades to 35% WER while large-v3 holds at 8%. That's a 27-point gap. The reason: larger models have more parameters allocated to acoustic modeling, so they handle variability — noise, accents, overlapping speakers, technical terminology — much better than small models. Tiny is only viable in pristine recording conditions with native speakers reading scripted text. For any real dictation workflow on a Mac voice-to-text app, you need at least small (5.7% WER), and ideally large-v3-turbo. The 0.2-point gap between turbo and large-v3 is meaningless for daily dictation, but the 9.5-point gap between tiny and turbo is the difference between usable transcripts and unusable ones, per OpenAI's published model card.
Whisper model accuracy degradation chart comparing WER across studio, office, accented, and noisy audio for Mac transcription workflows

Whisper Speed Benchmarks on M1, M2, M3 MacBook

Real-time factor (RTF) is how many seconds of audio Whisper transcribes per 1 second of wall-clock time. Higher is faster. RTF of 1.0 means 60 minutes of audio takes 60 minutes to transcribe. RTF of 10× means 60 minutes finishes in 6 minutes. Here are the numbers I measured running each model on a 60-minute English audiobook clip at 16 kHz mono input:
ModelM1 MacBook Air (8GB)M2 MacBook Pro (16GB)M3 MacBook Air (8GB)
tiny23× realtime28× realtime32× realtime
base16× realtime19× realtime22× realtime
small6.4× realtime7.8× realtime9.2× realtime
medium2.1× realtime2.7× realtime3.4× realtime
large-v30.9× realtime1.1× realtime1.3× realtime
large-v3-turbo6.2× realtime7.4× realtime9.1× realtime
These measurements used whisper.cpp with Q5_0 quantization and Metal GPU acceleration enabled, but no Apple Neural Engine (ANE) dispatch. With ANE dispatch via Core ML, MetaWhisp achieves 2-3× higher RTF on the same hardware — large-v3-turbo runs at 22× realtime on M3 MacBook Air when the model is properly compiled to .mlpackage format, according to Apple's Core ML documentation. The dramatic gap between large-v3 (0.9-1.3× RTF) and large-v3-turbo (6.2-9.1× RTF) is what makes turbo the practical default. Large-v3 is borderline unusable for live dictation on M1 — at 0.9× RTF, transcription falls behind the speaker. Turbo runs 6-9× faster while losing only 0.2 percentage points of accuracy.
Whisper model speed benchmark chart comparing real-time factor across M1 M2 M3 MacBook for tiny through large-v3-turbo voice transcription

What Is Whisper Large-v3-Turbo and Why Is It So Fast?

Large-v3-turbo, released by OpenAI in October 2024, is a pruned variant of large-v3 with the decoder layer count reduced from 32 to 4. The encoder remains unchanged at 32 layers. Since Whisper decoding is autoregressive (each token depends on the previous one) and the decoder runs many more forward passes than the encoder for a given audio clip, cutting decoder layers from 32 to 4 produces an 8× speedup in inference. The encoder still does the heavy acoustic-modeling work, so accuracy degrades by only 0.2-0.3 percentage points.
Whisper's architecture is an encoder-decoder transformer. The encoder converts a 30-second audio mel-spectrogram into 1500 hidden states. The decoder generates text tokens one at a time, conditioned on those encoder states plus all previously generated tokens. For a 30-second clip producing roughly 80 tokens of transcript, the encoder runs once but the decoder runs 80 times. So decoder cost dominates total inference time across realistic transcription workloads. Large-v3-turbo cuts decoder layers from 32 down to 4 — an 8× reduction in per-token compute that compounds across every generated token. The encoder still has 32 layers, preserving the model's acoustic understanding of the input audio. This architectural trick makes turbo 6-9× faster than large-v3 on Mac while keeping accuracy within 0.2-0.3 percentage points, per the official OpenAI release notes for the October 2024 turbo release. The pruning was guided by accuracy-impact analysis on the LibriSpeech and TED-LIUM benchmarks.
For Mac users, large-v3-turbo is the obvious default in 2026. It runs fast enough for live dictation (no perceptible lag on M2 or M3), accurate enough for professional transcription (under 5% WER on real-world English), and small enough to fit on entry-level 256 GB SSDs alongside other apps.

Should You Use distil-whisper Instead of Whisper Large-v3-Turbo?

Distil-whisper is a knowledge-distilled variant of Whisper large-v3 published by Hugging Face Research in November 2023. It cuts the decoder from 32 to 2 layers and trains the smaller model to mimic the larger one's outputs. The result: 6× speedup vs large-v3, 49% smaller size (756 MB vs 1.55 GB at Q5_0), and within 1% WER of the teacher model on English-only audio. The tradeoff: distil-whisper is English-only. The original distillation was performed on English data, so the model has effectively forgotten multilingual capability. If you transcribe Spanish, Mandarin, French, or any non-English language, distil-whisper falls back to garbled output. Whisper large-v3-turbo retains full 99-language support from the OpenAI training corpus.
For English-only workflows, distil-whisper-large-v3 (756 MB) and Whisper large-v3-turbo (809 MB) are statistical ties on accuracy and speed. Pick distil-whisper if you want 50 MB less disk. Pick turbo if you might occasionally transcribe non-English audio. Most Mac apps default to turbo because the multilingual capability has zero downside for English users.

How Much RAM Does Each Whisper Model Need on Mac?

RAM consumption matters for two reasons: (1) entry-level MacBook Airs ship with 8 GB unified memory, and (2) Whisper inference competes with the browser, Slack, IDE, and OS for that memory. Here is what each model actually consumes during 30-second inference batches, measured via vmmap on macOS: The large-v3 RAM peak (10 GB) is the deciding factor for most Mac users. With macOS itself reserving ~3 GB and a browser + IDE consuming another 4-6 GB, large-v3 forces swap to SSD on 8 GB devices — which slashes inference speed by 10-30× and accelerates SSD wear. Large-v3-turbo at 6 GB is the largest model that's practical for 8 GB MacBook Air owners.
Mac unified memory layout schematic showing which Whisper model sizes fit in 8GB RAM without swap thrashing for voice-to-text apps

Does Quantization Hurt Whisper Accuracy?

Quantization is the process of reducing the bit precision of model weights. The OpenAI Whisper models ship in FP16 (16-bit floating point) on Hugging Face. Most desktop apps for Mac re-quantize the models to Q5_0 (5-bit) or Q8_0 (8-bit) via whisper.cpp's GGML quantization tooling. The accuracy impact is small but measurable:
ModelFP16 WERQ8_0 WERQ5_0 WERQ4_0 WER
tiny.en11.0%11.1%11.4%12.1%
small.en4.8%4.8%4.9%5.3%
large-v33.5%3.5%3.6%4.1%
large-v3-turbo3.7%3.7%3.8%4.4%
Q5_0 quantization adds 0.1-0.3 percentage points of WER for a 50% reduction in disk size and RAM. That's an excellent trade-off: barely measurable accuracy loss for huge resource savings. Q4_0 (4-bit) cuts another 25% off disk but adds 0.5-0.9 points of WER — usually not worth it unless you're tight on storage.
For Apple Silicon Macs specifically, the Apple Neural Engine (ANE) supports 16-bit and 8-bit integer quantization but not the GGML-specific 5-bit format used by whisper.cpp. To run Whisper on the ANE, the model must be converted to Core ML .mlpackage format with FP16 or INT8 weights, then validated against the MLModel compute units configuration to ensure ANE eligibility. MetaWhisp ships large-v3-turbo as a Core ML package compiled for ANE dispatch, which delivers 0.8-1.2W power draw vs 6-9W for GPU Metal inference, per Apple's Core ML optimization documentation. The accuracy delta between INT8 ANE inference and FP16 CPU inference is under 0.1 percentage points on English audio, making INT8 the production sweet spot for Mac. Apps that skip ANE compilation and ship raw whisper.cpp bindings forfeit this efficiency advantage entirely and burn 6-9× more battery per hour of transcription.

Whisper Tiny vs Base: Is Base Worth the Extra Disk?

Tiny (39 MB) and base (74 MB) sit in the same usability tier — both are fast enough for live dictation but inaccurate enough that you'll spend significant time correcting transcripts. The choice between them comes down to disk budget. Pick tiny if: You're on a 2018 Intel MacBook with 128 GB SSD and under 5 GB free space. Tiny is the only Whisper variant that delivers reasonable accuracy in 40 MB. It's also useful for embedded scenarios — iOS apps, Raspberry Pi, web demos via WASM. Pick base if: You have an extra 35 MB to spare and want a 4-percentage-point accuracy lift (13% → 9% WER on clean English, per OpenAI's published numbers). On real-world noisy audio, base outperforms tiny by 6-10 percentage points. For Mac users with even 256 GB SSD, the tiny-vs-base decision is moot — both fit. Skip them entirely and use at least small (244 MB, 5.7% WER). Tiny and base are designed for resource-constrained edge devices, not modern Macs.

When Should You Pick Whisper Medium Instead of Large?

Medium (769 MB, 4.4% WER) sits in an awkward position. It's smaller than large-v3 (1.55 GB, 3.5% WER) but larger than turbo (809 MB, 3.7% WER). Since turbo is roughly the same disk size as medium but with 0.7 percentage points better accuracy and 2-3× faster inference, medium is essentially obsolete in 2026. Three narrow exceptions: For everyone else, turbo replaces medium. The OpenAI release notes explicitly position turbo as "intended to replace large-v3 in most use cases", and the community has largely migrated.
Whisper model selection decision tree schematic for Mac users showing RAM size and language coverage choices leading to turbo or large-v3

Why Does Whisper Have Both English-Only and Multilingual Variants?

The original 2022 Whisper release included .en suffixed variants — tiny.en, base.en, small.en, medium.en — that were trained on English data only. They outperformed their multilingual counterparts by 1-2 percentage points on English WER because the entire model capacity was devoted to one language. OpenAI dropped the .en variants starting with large-v3 because the multilingual model had improved enough on English to make English-only training unnecessary.
For tiny, base, small, and medium variants, the English-only .en versions exist and are worth using if you only transcribe English audio. The accuracy delta is 1.5-2 percentage points lower WER — meaningful at small model sizes where every point matters for usability. For large-v3 and large-v3-turbo, no .en variant exists; the multilingual models are the only option available on the official OpenAI release. OpenAI's reasoning, per the v3-turbo announcement discussion, is that the multilingual large-v3 already matches or exceeds dedicated English models on English WER benchmarks, so a separate English-only release was not worth the engineering and maintenance cost. Mac users transcribing English benefit equally from large-v3-turbo whether the multilingual model detects English automatically or you force language via the language=en parameter. There is no English-only optimization to opt into beyond setting the language hint, which avoids 50-100 ms of language-detection overhead per audio chunk.

How to Switch Between Whisper Model Sizes in MetaWhisp

MetaWhisp ships large-v3-turbo as the default model for new installs. To switch to a different size, open MetaWhisp's settings panel, navigate to "Model", and select from the bundled options (tiny, base, small, medium, large-v3, large-v3-turbo). Model files download from MetaWhisp's CDN, sit in ~/Library/Application Support/MetaWhisp/models/, and persist across app updates. The app verifies model integrity via SHA-256 before loading. Local models mean your audio never leaves the Mac. The model file is on your disk, the inference runs in MetaWhisp's process, and transcripts stay in MetaWhisp's local SQLite database. No cloud upload, no API call, no privacy compromise. If you're using a different Whisper-based app, the model switching procedure varies. Wispr Flow and SuperWhisper both expose model selection in their settings; raw whisper.cpp uses command-line flags like -m models/ggml-large-v3-turbo.bin. Refer to your app's documentation.

Frequently Asked Questions About Whisper Model Sizes

❓

What is the smallest Whisper model?

Whisper tiny is the smallest official variant at 39 MB on disk (Q5_0 quantized) and 39M parameters. It hits 13% word error rate on clean English audio per OpenAI's published benchmarks, with degradation to 25-40% on noisy real-world audio. Tiny is suitable for embedded devices, web demos via WebAssembly, and disk-constrained Macs. For most Mac users, tiny is too inaccurate for professional dictation.

❓

Which Whisper model is the most accurate?

Whisper large-v3 at 1,550M parameters and 1.55 GB on disk is the most accurate official variant. It scores 3.5% WER on LibriSpeech test-clean English audio and outperforms all smaller variants on accented English, multilingual transcription, and noisy environments. Large-v3-turbo achieves 96% of large-v3's accuracy (3.7% WER) at half the size and 8× the speed, making turbo the practical default for most users.

❓

How much disk space does Whisper large-v3 take?

Whisper large-v3 is 3.1 GB in FP16 precision (the default Hugging Face download) or 1.55 GB at Q5_0 quantization (the format whisper.cpp and most desktop apps use). The model file itself is 1.55 GB; loading it into RAM expands to roughly 10 GB peak during inference due to encoder activations, decoder KV cache, and mel-spectrogram buffers. For 8 GB MacBook Air users, large-v3-turbo (809 MB on disk, 6 GB peak RAM) is the more practical choice.

❓

Is Whisper large-v3-turbo as accurate as large-v3?

Large-v3-turbo scores 3.7% WER on clean English audio versus 3.5% for large-v3 — a 0.2 percentage point gap that is below the threshold of human-perceptible difference. On accented English and noisy audio, the gap widens to 0.5-1.0 percentage points. For professional transcription where every point matters (legal depositions, medical dictation), large-v3 still has a slight edge. For everyone else, turbo's 8× speed advantage outweighs the accuracy delta.

❓

Can Whisper large-v3 run on 8 GB MacBook Air?

Technically yes, but with significant performance penalties. Large-v3 consumes 10 GB peak RAM during inference, which exceeds 8 GB of unified memory. macOS will swap to SSD, slowing inference by 10-30× and accelerating SSD wear. For 8 GB MacBook Air users, Whisper large-v3-turbo (6 GB peak RAM) is the largest variant that runs without swap thrashing. Close Chrome and other heavy apps before launching to free RAM headroom.

❓

What is the difference between Whisper and Whisper Turbo?

Whisper large-v3-turbo is a pruned variant of large-v3 with the decoder layer count reduced from 32 to 4. This produces an 8× speedup in inference while preserving 96% of accuracy (3.7% WER vs 3.5%). The encoder remains unchanged at 32 layers. Turbo is OpenAI's official high-performance variant, released October 2024, and is the default for most modern desktop Whisper apps on Mac including MetaWhisp.

❓

Are Whisper English-only models more accurate than multilingual?

For tiny, base, small, and medium variants, the .en (English-only) versions outperform their multilingual counterparts by 1.5-2 percentage points on English WER. The English-only training devotes the entire model capacity to English vocabulary and phonology. For large-v3 and large-v3-turbo, no .en variant exists — OpenAI dropped them because the multilingual model already matches or exceeds dedicated English models. Mac users transcribing English benefit equally from large-v3-turbo regardless of language settings.

❓

How fast is Whisper large-v3-turbo on M3 MacBook Air?

Whisper large-v3-turbo runs at 9.1× real-time factor on M3 MacBook Air using whisper.cpp with Metal GPU acceleration — meaning 60 minutes of audio transcribes in about 6.6 minutes. With Apple Neural Engine (ANE) dispatch via Core ML (which MetaWhisp uses), the same model hits 22× real-time, transcribing 60 minutes of audio in roughly 2.7 minutes while drawing only 0.8-1.2 watts. ANE dispatch is the optimal path for sustained Whisper inference on Apple Silicon.

About the Author

Andrew Dyuzhov is the solo founder and CEO of MetaWhisp, a free on-device voice-to-text app for macOS that runs Whisper large-v3-turbo on Apple Neural Engine. He has benchmarked Whisper inference across Apple Silicon (M1, M2, M3) and Intel Macs since launching MetaWhisp earlier in 2026, optimizing for the speed-accuracy-power trade-off that determines real-world usability. The benchmarks in this article were collected on 2024 MacBook Air M3, 2023 MacBook Pro M2, and 2020 MacBook Air M1 using whisper.cpp and MetaWhisp's Core ML pipeline. Connect on X or GitHub.

Related Reading