Qwen3-ASR-0.6B — CoreML
CoreML conversion of Qwen/Qwen3-ASR-0.6B for Apple Neural Engine.
Contains both audio encoder and text decoder for full Neural Engine inference (no GPU required).
Models
| Model | Description | Quantization |
|---|---|---|
encoder.mlmodelc |
Audio encoder (mel → embeddings) | INT8 palettized |
embedding.mlmodelc |
Token embedding lookup | INT8 palettized |
decoder.mlmodelc |
Text decoder with KV cache (28 layers) | INT8 palettized |
encoder_int4.mlpackage |
Audio encoder source | INT4 palettized |
encoder_int8.mlpackage |
Audio encoder source | INT8 palettized |
Usage
Full CoreML pipeline (encoder + decoder on Neural Engine):
[1;38;5;196mWelcome to Swift![0m
[1mSubcommands:[0m
[1mswift build[0m Build Swift packages [1mswift package[0m Create and work on packages [1mswift run[0m Run a program from a package [1mswift test[0m Run package tests [1mswift repl[0m Experiment with Swift code interactively
Use [1mswift --version[0m for Swift version information.
Use [1mswift --help[0m for descriptions of available options and flags.
Use [1mswift help <subcommand>[0m for more information about a subcommand.
Hybrid mode (CoreML encoder + MLX decoder on GPU):
Architecture
- Audio encoder: 18-layer Whisper-style transformer (896 dim, 14 heads)
- Text decoder: 28 layers, 1024 hidden, 16 heads (8 KV heads)
- KV cache: Fixed 1024 tokens via CoreML MLState
- Requires: macOS 15+ / iOS 18+ (full CoreML mode)
Links
- Swift library: soniqo/speech-swift
- Base model: Qwen/Qwen3-ASR-0.6B
- Guide: soniqo.audio/guides/transcribe
- Docs: soniqo.audio
- GitHub: soniqo/speech-swift