software-mansion/react-native-executorch-qwen-3-asr

qwen-3-asr

This repository hosts Qwen3-ASR exported for the React Native ExecuTorch library as ExecuTorch .pte programs, ready to run on device.

Variants

Path	Backend	Precision	Audio per pass
0_6b/xnnpack/qwen3_asr_0_6b_xnnpack_8da4w.pte	xnnpack	8da4w	30 s
0_6b/mlx/qwen3_asr_0_6b_coreml_mlx_4w.pte	Core ML (audio tower) + MLX	fp16 tower, 4w decoder	30 s

Decoder linears are 8da4w (group 32) except decoder layers 0–1, which stay at 8da8w; the audio tower is 8da8w with an int8 conv stem. WER on the first 100 LibriSpeech test-clean clips: 2.46% (fp32: 2.14%).

The Core ML + MLX file runs the audio tower on Core ML in fp16 and the decoder on MLX with 4-bit weights (group 32, layers 0–1 at 8 bits): WER 2.51%. It needs both backends in the runtime.

Repository structure

0_6b/tokenizer.json                              10.9 MiB
0_6b/xnnpack/config.json
0_6b/mlx/config.json
0_6b/mlx/qwen3_asr_0_6b_coreml_mlx_4w.pte      771 MiB
0_6b/xnnpack/qwen3_asr_0_6b_xnnpack_8da4w.pte   675 MiB

Compatibility

These files are published for the ExecuTorch v1.4.1 runtime. ExecuTorch gives no forward compatibility guarantee, so an older runtime may fail to load them.

To use them in React Native ExecuTorch, pass the model constant shipped in the library's model registry to the corresponding task pipeline. See the documentation.

To load these files in your own ExecuTorch runtime, read the compatibility note first.