moona3k/mlx-qwen3-asr-0.6b-4bit

moona3k/mlx-qwen3-asr-0.6b-4bit

4-bit (group size 64) MLX quantization of Qwen/Qwen3-ASR-0.6B for mlx-qwen3-asr, a ground-up MLX reimplementation of Qwen3-ASR for Apple Silicon. No PyTorch, no transformers, no model conversion step on the user's side.

Usage

pip install -U mlx-qwen3-asr
mlx-qwen3-asr audio.wav --model moona3k/mlx-qwen3-asr-0.6b-4bit

import mlx_qwen3_asr as m

result = m.transcribe("audio.wav", model="moona3k/mlx-qwen3-asr-0.6b-4bit")
print(result.text)

Requires mlx-qwen3-asr >= 0.4.3. Mixed encoder/decoder widths need the per-module loader introduced in 0.4.3; older versions fail to load this artifact with a shape error. Word timestamps (--timestamps) and the HTTP server (mlx-qwen3-asr serve --model moona3k/mlx-qwen3-asr-0.6b-4bit) work unchanged.

What is quantized

• —Text decoder Linear and Embedding layers are affine-quantized to 4 bits; the audio encoder is quantized to 8 bits (group size 64 for both). The encoder carries most of the 4-bit quality loss: on 0.6B, all-4-bit scored 2.63% WER against 2.37% with an 8-bit encoder (fp16 2.33%), for 90 MB more.
• —Remaining floating tensors (scales, biases, norms, conv stem) are stored in float16, so inference runs in float16 end to end.
• —lm_head is tied to the token embedding in the source model and is not stored twice; the loader re-ties it.
• —Download size: 517M (fp16 source: 1.8 GB).

Quality

LibriSpeech test-clean, 100 speaker-balanced clips (speaker_round_robin), greedy decoding, Apple M4 Pro, MLX 0.30.6:

Hypotheses: 17 of 100 hypotheses differ from fp16, mostly proper-noun spellings and British/American variants; WER is within 0.05pp of fp16.

Latency envelope from the committed quantization matrix (docs/benchmarks/2026-09-07-quant-matrix-test-clean-speaker100.md, 0.6B): 8-bit runs about 2.4x and 4-bit about 2.7x faster than fp16 on a 10 s clip. Per-sample JSON for this artifact's evaluation is committed in the mlx-qwen3-asr repository as docs/benchmarks/2026-09-19-quantized-artifacts-librispeech-test-clean-100-0.6B_4bit.json.

Reproduce

git clone https://github.com/moona3k/mlx-qwen3-asr && cd mlx-qwen3-asr && git checkout v0.4.3
python scripts/convert.py --model Qwen/Qwen3-ASR-0.6B --quantize 4 --encoder-bits 8 \
  --group-size 64 --dtype float16 --output-dir Qwen3-ASR-0.6B-4bit-g64
python scripts/eval_librispeech.py --model Qwen3-ASR-0.6B-4bit-g64 --samples 100 --sampling speaker_round_robin

License

Apache-2.0, following the source model.