Qwen3-ASR-1.7B — MLX (int8)
MLX-native int8 conversion of Qwen3-ASR-1.7B for local automatic speech recognition on Apple Silicon — English, Chinese, and mixed Chinese/English. It runs through the mlx-speech runtime with no PyTorch and no cloud API at inference time. Weights ship as plain .safetensors.
This is the default Qwen3-ASR build in mlx-speech. In local Apple Silicon checks the int8 build produced transcripts identical to the bf16 build on our English test clips. For the unquantized reference, see the bf16 build.
Model Details
- Developed by: App Automaton
- Upstream model:
Qwen/Qwen3-ASR-1.7B (code: QwenLM/Qwen3-ASR)
- Task: automatic speech recognition — English, Chinese, and mixed Chinese/English (upstream supports more languages; the
mlx-speech v0 path validates en/zh/mixed)
- Precision: int8 affine, group_size 64, applied to the Linear and Embedding layers across the audio tower and text decoder; Conv2D layers are kept unquantized
- Runtime: MLX on Apple Silicon
- Input: 16 kHz mono audio
Variants
| Repo |
Precision |
Notes |
qwen3-asr-1.7b-bf16-mlx |
bf16 |
unquantized reference |
qwen3-asr-1.7b-int8-mlx (this repo) |
int8 affine |
default — recommended for most users |
How to Get Started
Install mlx-speech, then load by alias (downloads on first use):
import mlx_speech
asr = mlx_speech.asr.load("qwen3-asr-1.7b-int8")
result = asr.generate("speech.wav") # language=None auto-detects
print(result.language, result.text)
# For mixed Chinese/English, forcing Chinese preserves Chinese characters best:
result = asr.generate("mixed-speech.wav", language="Chinese")
mlx-speech asr --model qwen3-asr-1.7b-int8 --audio speech.wav
Or download the weights once and load by local path:
hf download appautomaton/qwen3-asr-1.7b-int8-mlx \
--local-dir models/qwen3_asr_1_7b/mlx-int8
asr = mlx_speech.asr.load("models/qwen3_asr_1_7b/mlx-int8")
Notes
- This repo contains the quantized MLX runtime artifact only.
- The quantization mode (
affine, group_size 64, 8-bit) is stored in config.json and re-applied automatically on load, so it cannot desync from the weights.
- v0 is single-pass offline ASR; streaming, timestamps, and long-audio chunking are deferred.
- Accuracy was checked relative to the bf16 build on English clips; broader Chinese/mixed evaluation is ongoing.
Links
License
Apache 2.0, following the upstream Qwen3-ASR release. Refer to the original Qwen release for current terms.
