Glimpse-Dictation/Whisper-Large-V3-coreml

Whisper Large V3: Core ML encoder and GGUF decoders

This repository contains the optional Apple Silicon package for Whisper Large V3 in Glimpse: a compiled Core ML encoder and decoder-only GGUFs. A decoder-only GGUF carries the decoder, tokenizer and metadata without the encoder weights, so the encoder is not downloaded twice. The encoder also works beside a full Whisper GGUF or whisper.cpp ggml model. Glimpse uses the encoder for every Whisper Large V3 quantization and for Distil-Whisper Large V3.5.

File	Download size	Purpose
whisper-large-v3-encoder.mlmodelc.zip	1.18 GB	Compiled Core ML encoder
whisper-large-v3-Q8_0-decoder.gguf	969.0 MB	Decoder, tokenizer and metadata of whisper-large-v3-Q8_0.gguf

A decoder-only GGUF needs the encoder. On Windows, Intel Macs, or anywhere without Core ML, use the full model file instead. The Core ML package requires Apple Silicon.

Provenance

The model originates from OpenAI Whisper Large V3, licensed Apache 2.0. The encoder was converted from whisper.cpp's F16 ggml-large-v3.bin with scripts/convert-whisper-gguf-to-coreml.py from our transcribe.cpp fork, then zipped with ditto -c -k --keepParent. Each decoder-only GGUF comes from the full GGUF listed below via scripts/extract-coreml-decoder.py, which copies the decoder and frontend tensors and metadata byte for byte, omits the encoder tensors and sets stt.whisper.decoder_only=true.

Artifact	SHA-256
Encoder ZIP	504d9b03b34c689e91d61fb97d4c97eca979fc4d21e65d9c2963b7481cbc66dc
Original whisper-large-v3-Q8_0.gguf	2fa1a5f179f8a511a53e2108db270aa4af3ce08cd976af4180e2854666bb4ba3
whisper-large-v3-Q8_0-decoder.gguf	a40e809cfbb958dc9053609ff9e54595cd011e960bcaebfa0f281ae93995ebe4

Runtime requirements

Extract the ZIP next to the model file, keeping the whisper-large-v3-encoder.mlmodelc directory name. Glimpse-Speech looks for the companion beside the model. A full model file falls back to its own encoder when the companion isn't there; a decoder-only GGUF has no encoder to fall back to, so a session fails without the companion. Decoder-only GGUFs need a transcribe.cpp build with decoder-only Whisper support.

The encoder follows Apple's Neural Engine transformer layout and computes in FP16. It takes one 30-second Whisper window. Large encoders are split into chained Core ML programs, because the Neural Engine compiler rejects a single program of that size and falls back to the CPU.

Core ML may run some operations on the CPU; this is not a guarantee of exclusive Neural Engine execution.

Glimpse

This encoder speeds up on-device dictation on Apple Silicon in Glimpse, a free, open-source dictation app for Mac and Windows. The source is on GitHub.