doggy8088/Qwen3-ASR-1.7B-MLX-4bit

This repo contains an MLX conversion of Qwen/Qwen3-ASR-1.7B.

Variant

• Format: MLX safetensors
• Weights dtype: bfloat16
• Quantization: 4-bit
• Custom architecture helper: qwen3_asr_mlx.py

Files

• model*.safetensors: converted MLX weights
• config.json: converted model config
• preprocessor_config.json, tokenizer_config.json, vocab.json, merges.txt: tokenizer / processor assets copied from the source model
• qwen3_asr_mlx.py: custom MLX architecture + processor helper used for this conversion

Notes

• Source model: Qwen/Qwen3-ASR-1.7B
• Local conversion directory: Qwen3-ASR-1.7B-MLX-4bit
• This is an Apple Silicon / MLX-oriented conversion.
• For quantized variants, quantizable linear layers are quantized while non-quantizable layers remain in floating point.