Argmax's WhisperKit Beats OpenAI and Deepgram With 11M Downloads on Apple Silicon
Argmax's on-device Whisper for Apple Silicon crossed 11M downloads, hit v1.0, and beats cloud ASR APIs on both latency and accuracy.
PRO- WhisperKit CoreML crossed 11M downloads on Hugging Face, first Apple Silicon-only model to hit that scale
- ICML 2025 paper shows 0.46s latency and 2.2% WER, beating OpenAI, Deepgram, and Fireworks
- New OD-MBP compression shrinks Whisper Large v3 Turbo from 1.6 GB to 0.6 GB with under 1% WER loss
- Stateful Core ML models cut decoder energy per pass by 75% (1.5W to 0.3W) on M3 Neural Engine
- Repo renamed to argmax-oss-swift, bundling WhisperKit, SpeakerKit, and TTSKit with Swift 6 support
- Install with
brew install whisperkit-clior add the Swift package in Xcode, MIT license
A pure-Swift speech recognition package running Whisper entirely on Apple Silicon has quietly become one of the most-downloaded models on Hugging Face, and its authors just published benchmarks showing it beats OpenAI, Deepgram, and Fireworks on real-time transcription. WhisperKit CoreML from Argmax now sits at over 11 million all-time downloads, and the underlying framework has graduated from a single kit into a full open-source SDK that bundles speech-to-text, speaker diarization, and text-to-speech in one MIT-licensed Swift package.
Argmax reports that WhisperKit is the first Apple Silicon-only model to cross 10 million monthly downloads on Hugging Face, with usage growing roughly tenfold in about 100 days. That trajectory, combined with a fresh ICML paper and a v1.0 SDK release, makes this worth a closer look for anyone shipping voice features on Mac or iOS.
Beating cloud APIs on their own turf
The headline claim from the WhisperKit paper: an on-device system that matches or exceeds server-side APIs on both speed and word error rate. The team benchmarked against a frontier model (OpenAI gpt-4o-transcribe), a proprietary ASR service (Deepgram nova-3), and a hosted open-source model (Fireworks large-v3-turbo).
WhisperKit ties for lowest latency at 0.46s while posting the best accuracy at 2.2% WER on TIMIT. Fireworks matched raw latency but issued an order of magnitude more corrections, which the authors argue degrades user experience even when average speed looks fast. OpenAI's gpt-4o-transcribe doesn't support hypothesis text streams at all, so its confirmed-text latency lags far behind.
How a 1B-parameter Transformer fits on your phone
Getting Whisper Large v3 Turbo, a 1-billion-parameter encoder-decoder, to run in real time on a phone required rethinking three things: the streaming architecture, the compression scheme, and how the model talks to the Apple Neural Engine.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.