Qwen3-ASR-1.7B — MLX 4-bit
MLX 4-bit quantized conversion of Qwen/Qwen3-ASR-1.7B for Apple Silicon inference.
Model Details
| Detail | Value |
|---|---|
| Architecture | Whisper-style audio encoder + Qwen3 text decoder |
| Parameters | 1.7B |
| Quantization | 4-bit (group_size=64, text decoder only) |
| Audio encoder | float16 (24 layers, 1024 dim, 16 heads) |
| Size | ~2.1 GB |
| Languages | Multilingual (EN, ZH, JA, KO, FR, DE, ES, and more) |
Usage
let model = try await Qwen3ASRModel.fromPretrained(
modelId: "aufklarer/Qwen3-ASR-1.7B-MLX-4bit"
)
let text = model.transcribe(audio: samples, sampleRate: 16000)
audio transcribe audio.wav --model aufklarer/Qwen3-ASR-1.7B-MLX-4bit
Variants
| Variant | Size | Model ID |
|---|---|---|
| 4-bit | ~2.1 GB | aufklarer/Qwen3-ASR-1.7B-MLX-4bit |
| 8-bit | ~3.2 GB | aufklarer/Qwen3-ASR-1.7B-MLX-8bit |
| 0.6B 4-bit | ~680 MB | aufklarer/Qwen3-ASR-0.6B-MLX-4bit |
| 0.6B 8-bit | ~1.0 GB | aufklarer/Qwen3-ASR-0.6B-MLX-8bit |
Links
- Swift library: soniqo/speech-swift
- Base model: Qwen/Qwen3-ASR-1.7B
- Guide: soniqo.audio/guides/transcribe
- Docs: soniqo.audio
- GitHub: soniqo/speech-swift