FluidInference/qwen3-asr-0.6b-coreml

Qwen3-ASR 0.6B CoreML

Core ML conversion of Qwen/Qwen3-ASR-0.6B for on-device speech recognition on Apple platforms (iOS/macOS).

Model Variants

Variant	Size	Description
f32/	~2.5 GB	Full precision (Float32) - highest accuracy
int8/	~0.7 GB	Quantized (Int8) - smaller, faster

Features

• 30+ languages including English, Chinese, Japanese, Korean, and more
• On-device inference - no internet required
• Autoregressive decoder with KV-cache support
• Processes audio in 1-second chunks (100 mel frames)

Benchmarks (M4 Pro)

Dataset	WER	CER	RTFx
LibriSpeech test-clean (2620 files)	4.4%	1.9%	2.8x
AISHELL-1 test (100 files)	4.6%	3.7%	4.5x

Official PyTorch model: 2.11% WER on LibriSpeech test-clean

Usage with FluidAudio

import FluidAudio

let manager = Qwen3AsrManager()
try await manager.loadModels()

let samples = try AudioConverter().resampleAudioFile(path: "audio.wav")
let transcript = try await manager.transcribe(
    audioSamples: samples,
    language: "en",
    maxNewTokens: 512
)
print(transcript)

Model Architecture

• Encoder: Audio encoder (Whisper-style mel spectrogram input)
• Decoder: 28-layer transformer decoder with 1024 hidden size
• Tokenizer: Qwen tokenizer with special ASR tokens

License

Apache 2.0 - Same as the original Qwen3-ASR model.

Credits

• Original model: https://huggingface.co/Qwen/Qwen3-ASR-0.6B by Alibaba Qwen Team
• Paper: https://arxiv.org/abs/2601.21337
• CoreML conversion: https://github.com/FluidInference/FluidAudio

Citation

@article{qwen3asr, title={Qwen3-ASR Technical Report}, author={Qwen Team}, journal={arXiv preprint arXiv:2601.21337}, year={2025} }

For the HuggingFace metadata UI, fill in:

• License: Apache 2.0
• Base model: Qwen/Qwen3-ASR-0.6B
• Pipeline: automatic-speech-recognition
• Library: coreml
• Languages: en, zh, ja, ko, + others