mlx-qwen3-asr
A from-scratch reimplementation of the Qwen3-ASR speech recognition model for Apple's MLX framework, running natively on Apple Silicon GPUs without PyTorch.
mlx-qwen3-asr ports Qwen3-ASR, an open speech recognition model that exceeds Whisper-large-v3 on several languages and 22 Chinese dialects, to Apple's MLX so it runs natively on M1 through M4 Macs via Metal, with no PyTorch dependency for core transcription. Every layer, including the audio encoder, the interleaved MRoPE text decoder and the mel-spectrogram frontend, is reimplemented rather than wrapped.
It supports both the 0.6B and 1.7B model sizes, long audio split into chunks with no truncation, word-level timestamps through a native forced aligner, optional speaker diarization via pyannote, 4-bit and 8-bit quantization, and output as text, JSON, SRT, VTT or TSV. A built-in HTTP server exposes the pipeline with OpenAI-API compatibility and Bearer token auth, and it also supports experimental speculative decoding and streaming transcription.