CoolFace
Modelpublic

aufklarer/Qwen3-ASR-CoreML

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face
4likes755downloads
Model Card

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

ModelDescriptionQuantization
encoder.mlmodelcAudio encoder (mel → embeddings)INT8 palettized
embedding.mlmodelcToken embedding lookupINT8 palettized
decoder.mlmodelcText decoder with KV cache (28 layers)INT8 palettized
encoder_int4.mlpackageAudio encoder sourceINT4 palettized
encoder_int8.mlpackageAudio encoder sourceINT8 palettized

Usage

Full CoreML pipeline (encoder + decoder on Neural Engine):

Welcome to Swift!

Subcommands:

swift build Build Swift packages swift package Create and work on packages swift run Run a program from a package swift test Run package tests swift repl Experiment with Swift code interactively

Use swift --version for Swift version information.

Use swift --help for descriptions of available options and flags.

Use swift help <subcommand> 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