Can Whisper Run Offline on Mac? Yes, Here's How (2026)
Yes, Whisper runs fully offline on Mac after a one-time model download. Hardware requirements, three setup paths, and M1-M4 benchmarks inside.
Sonicribe Team
Product Team
Table of Contents
Can Whisper Run Offline on Mac?
Yes. OpenAI's Whisper runs 100% offline on a Mac after a one-time model download. The model weights are pulled from the internet once, saved to disk, and every transcription after that is computed locally on the CPU, GPU, or Apple Neural Engine. No audio, no text, and no metadata leaves the device during inference.
This guide explains what "offline" actually means for Whisper, the hardware required for each model size, the three common ways to run it on macOS, and real-world benchmarks on Apple Silicon from M1 through M4.
Short Answer
Whisper is fully offline-capable on Mac. You need an internet connection once to download model weights (between 39 MB and 1.55 GB depending on size). After that first download, inference runs entirely on-device using Core ML, Metal, or the Apple Neural Engine. You can verify zero network activity with Little Snitch or Apple's built-in firewall. A MacBook Air M1 with 8 GB of RAM can run Whisper small comfortably. A MacBook Pro M2 Max or newer runs Whisper large-v3 faster than real time.
What "Offline" Actually Means for Whisper
Two phases matter here, and confusing them is where most "does Whisper need internet" questions come from.
Phase 1, model download. The first time you run Whisper, the chosen model file (for example,ggml-base.en.bin or large-v3.pt) is fetched from Hugging Face or OpenAI's CDN. This happens exactly once per model. The file is cached on disk, typically under ~/.cache/whisper/ or inside the app bundle.
Phase 2, inference. Every transcription after that runs against the cached weights. No request is sent to any server. The audio never leaves the machine. The transcript is produced locally.
If the question is "does Whisper need internet to transcribe my audio," the answer is no. If the question is "does Whisper need internet the very first time I use a new model size," the answer is yes, briefly.
You can pre-download model files on a connected machine and copy them to an air-gapped Mac. Whisper does not phone home to validate weights.
Hardware Requirements by Model Size
Whisper ships in five multilingual sizes plus English-only variants. The table below pairs each size with the minimum Apple Silicon hardware that runs it at or above real-time speed.
| Model | Disk size | RAM needed | Minimum Mac | Best for |
|---|---|---|---|---|
| tiny | 39 MB | ~1 GB | Any M-series | Quick notes, low-latency dictation |
| base | 74 MB | ~1 GB | Any M-series | Everyday dictation, short audio |
| small | 244 MB | ~2 GB | M1 / M1 Air (8 GB) | Balanced accuracy and speed |
| medium | 769 MB | ~5 GB | M1 Pro / M2 (16 GB) | Meetings, interviews, podcasts |
| large-v3 | 1.55 GB | ~10 GB | M2 Max / M3 Pro | Highest accuracy, multilingual |
The English-only variants (tiny.en, base.en, small.en, medium.en) use the same disk and RAM but deliver slightly better accuracy on English audio. There is no large.en, large-v3 is already multilingual by design.
Quantized versions (q5_0, q8_0, int8) cut RAM by 30-50% at a small accuracy cost. whisper.cpp supports quantization natively. Most GUI apps expose quantized variants as "fast" or "optimized" model options.
Three Ways to Run Whisper Offline on Mac
There are three practical paths, each suited to a different user.
1. whisper.cpp (command line, for developers)
whisper.cpp is Georgi Gerganov's C/C++ port of Whisper with Core ML and Metal acceleration. It is the fastest offline Whisper implementation on Apple Silicon and the reference for most GUI apps.
Install via Homebrew:
brew install whisper-cpp
Download a model:
bash ./models/download-ggml-model.sh base.en
Transcribe a file:
whisper-cpp -m models/ggml-base.en.bin -f audio.wav
whisper.cpp is ideal if you want scripted batch transcription, want to integrate Whisper into a pipeline, or need quantized models on lower-end Macs. The learning curve is real, and there is no GUI or live dictation.
2. MacWhisper (file-based GUI, for casual users)
MacWhisper is a native macOS app that wraps whisper.cpp in a simple interface. You drag an audio or video file in, pick a model, and get a transcript. It supports most common audio formats and exports to SRT, VTT, and TXT.
MacWhisper is the right pick if you have a pile of recordings to transcribe (lectures, interviews, voice memos) and do not need live dictation into other apps. It has a free tier with base models and a paid tier for large-v3.
3. Sonicribe (prebuilt app with hotkey and live dictation)
Sonicribe is a macOS dictation app built on Whisper with a global hotkey, auto-paste into any app, 99+ language support, and 10 vocabulary packs. The free tier includes 10,000 words per week. The paid tier is a one-time $79 purchase.Sonicribe is for users who want a prebuilt dictation workflow without configuring command-line tools. You press a hotkey, speak, and the transcript appears in whatever app has focus (Slack, VS Code, Gmail, Notes). The underlying engine is the same Whisper family of models, running locally on Core ML.
Sonicribe is one of several good options, not the only one. If you prefer open source and do not need live dictation, whisper.cpp or MacWhisper are strong alternatives.
For a broader look at offline speech-to-text tools, see our offline speech-to-text comparison.
Performance Benchmarks on M1 through M4
Benchmarks are reported as real-time factor (RTF), the ratio of audio duration to transcription time. An RTF of 0.3 means 1 minute of audio transcribes in 18 seconds. Lower is faster. Numbers below use whisper.cpp with Core ML acceleration on English audio, 16 kHz mono WAV input.
MacBook Air M1 (8 GB RAM)- tiny.en: RTF ~0.08 (12x faster than realtime)
- base.en: RTF ~0.12
- small.en: RTF ~0.25
- medium.en: swappy, usable with care
- large-v3: not recommended on 8 GB
- base.en: RTF ~0.08
- small.en: RTF ~0.15
- medium.en: RTF ~0.35
- large-v3: RTF ~0.7 (slower than realtime, but acceptable for batch)
- small.en: RTF ~0.08
- medium.en: RTF ~0.18
- large-v3: RTF ~0.3 (faster than realtime)
- medium.en: RTF ~0.14
- large-v3: RTF ~0.25
- large-v3: RTF ~0.15 (about 6x realtime)
- large-v3-turbo: RTF ~0.08
The Apple Neural Engine accelerates Core ML-encoded encoder layers. The decoder still runs on CPU or GPU via Metal. Sonicribe, MacWhisper, and whisper.cpp all take advantage of this split when Core ML weights are available.
Numbers above are approximate and depend on audio quality, language, and whether you use quantization. Your mileage will vary by ±20%.
Common Gotchas
First-run download stalls. The model file is downloaded on first use, not at install. A slow connection can make the app appear frozen. Check disk I/O or network activity before assuming a crash. Disk space. Large-v3 is 1.55 GB. If you also keep medium and small, plan for around 2.5 GB of model cache. Some GUIs cache models inside the app bundle, which complicates deletion. RAM pressure on 8 GB Macs. Medium and large-v3 will cause heavy swap on base M1 and M2 Air. Stick to small or smaller on 8 GB machines, or use a quantized medium. Audio format conversion. whisper.cpp wants 16 kHz mono WAV. If you feed it MP3 or stereo, it converts via ffmpeg, which must be installed separately. MacWhisper and Sonicribe handle conversion internally. Core ML weights are separate. whisper.cpp gets its Apple Silicon boost from Core ML-encoded weights (ggml-base.en-encoder.mlmodelc), which must be generated or downloaded separately. Prebuilt apps bundle them.
Language detection adds latency. If you do not specify a language, Whisper runs a language detection pass first. Pinning the language to English (or whichever you speak) shaves a few hundred milliseconds off short clips.
Privacy Implications: Verified Offline
After the first model download, an offline Whisper setup makes zero outbound network calls for transcription. You can verify this three ways.
Little Snitch. Enable "Alert Mode" and run a transcription. No rule prompts means no connections attempted. macOS Application Firewall. Block all incoming connections, turn on stealth mode, and transcribe. Whisper still works. Airplane mode. Turn Wi-Fi off. Run a transcription. If it completes, inference is local.Sonicribe publishes a zero-telemetry policy for its transcription path, meaning audio and transcript data never leave the device. Three categories of traffic do remain networked in the default setup, and it is worth being explicit about them:
1. Model downloads on first use of a new model size.
2. License check on install and occasional revalidation for paid licenses.
3. Usage counter for the free tier, to track the 10,000 words per week cap. This counter reports a word count only, not audio or transcript content.
Users who want an air-gapped install can disable the license check and free tier counter after activating a paid license. whisper.cpp and MacWhisper have no license check, and whisper.cpp has no counters at all.
For a broader privacy comparison across free speech-to-text tools, see our best free speech-to-text guide.
Why Apple Silicon Makes This Practical
Local Whisper was usable on Intel Macs, but slow. Apple Silicon changed three things:
Unified memory. The CPU, GPU, and Neural Engine share the same RAM pool. No cross-device copies during inference. Neural Engine (ANE). 16-18 TOPS of dedicated matrix throughput on M1-M3, and 38 TOPS on M4 Max. Whisper's encoder maps well to ANE via Core ML. Metal Performance Shaders. The decoder runs on GPU through MPS, which is faster than CPU for beam search and autoregressive generation.The combined effect is that a 2024 MacBook Air runs whisper-small faster than a 2020 Intel MacBook Pro ran whisper-tiny. Batch transcription of a 1-hour interview on M3 or M4 with large-v3 takes about 15-25 minutes, fast enough to be practical for daily work.
Frequently Asked Questions
Does Whisper need an internet connection?
Whisper needs internet once, to download model weights from OpenAI or Hugging Face. After that, every transcription runs fully offline on your Mac. You can verify this by disabling Wi-Fi and running a transcription, or by monitoring network activity with Little Snitch or macOS's built-in firewall. For air-gapped environments, you can download weights on a connected machine and copy them over.
Can I run Whisper on a MacBook Air M1?
Yes. A MacBook Air M1 with 8 GB of RAM comfortably runs Whisper up to the small model (244 MB), which delivers 92-95% accuracy on clean English audio at about 4x real-time speed. Medium and large-v3 are technically runnable but cause heavy memory swap on 8 GB machines. For the best balance of accuracy and speed on base M1/M2 Air hardware, stick to small or base models, or use a quantized medium.
How much disk space does offline Whisper use?
A single Whisper model takes between 39 MB (tiny) and 1.55 GB (large-v3). If you install multiple sizes, plan for around 2.5 GB total. Core ML encoder weights and quantized variants add another 500 MB to 1 GB. Most Whisper apps cache models in ~/.cache/whisper/, ~/Library/Application Support/, or inside the app bundle, depending on the implementation.
Is Whisper fully private when run locally?
Yes. Local Whisper transcription sends zero audio, transcript, or metadata over the network after the initial model download. You can confirm this with airplane mode, Little Snitch, or the macOS Application Firewall. Prebuilt apps may still make unrelated network calls for license validation or usage counting, which are separate from the transcription pipeline. whisper.cpp, the open source reference implementation, makes no network calls at all after model download.
What's the easiest way to run Whisper offline on Mac?
For non-technical users, Sonicribe or MacWhisper are the easiest paths. Sonicribe provides a global hotkey and auto-paste into any app for live dictation. MacWhisper is drag-and-drop for audio files. Both are prebuilt, signed, and run locally on Whisper with Core ML acceleration. For developers who want to script transcription, whisper.cpp via Homebrew is the fastest and most flexible option, and it has no license or account requirements.
The Bottom Line
Whisper runs fully offline on Mac. The only internet requirement is a one-time model download, between 39 MB and 1.55 GB. Every transcription after that runs on-device using Core ML, Metal, and the Apple Neural Engine. A MacBook Air M1 handles Whisper small comfortably. A MacBook Pro M2 Max or newer runs large-v3 faster than real time.
Three setup paths cover most needs. whisper.cpp for developers who want command-line control. MacWhisper for users with piles of audio files to batch-transcribe. Sonicribe for users who want live dictation with a hotkey, auto-paste into any app, and a free tier that covers most daily use.
All three process audio locally. All three work with Wi-Fi off. Pick based on how you work, not on whether offline Whisper is possible, it already is.
Want to try offline Whisper without setting up a toolchain? Download Sonicribe, 10,000 words per week free, no account required, audio stays on your Mac.
Related Reading
Find a workflow for your next draft.
Explore voice-to-text examples, custom formatting, and the new cloud plans. Sonicribe requires a verified Google account and an internet connection. The separate new app download is being prepared.
New app release status
