Peacockery/MOSS-Transcribe-preview-2B-MLX

MOSS-Transcribe-preview-2B — MLX conversion

MLX-weight conversion of OpenMOSS-Team/MOSS-Transcribe-preview-2B, an English speech-to-text model pairing a Qwen3-1.7B-base decoder with a Qwen3-Omni-MoE audio encoder and a gated-MLP adapter (~2.4B parameters, Apache-2.0 upstream).

Converted from the upstream bf16 safetensors by the peacock-asr moss-mlx-conversion project.

Variants

Directory	Precision	Size	Notes
MOSS-Transcribe-preview-2B-bf16/	bf16	4.6 GB	Reference MLX-layout conversion; matched PyTorch transcript exactly on the LibriSpeech smoke fixture.
MOSS-Transcribe-preview-2B-text-decoder-4bit-g64/	4-bit affine, group 64 (text decoder only)	2.7 GB	Strongest quantized candidate in the conversion project: 2.48 RTFx on the first 20 LibriSpeech clean-test rows with no bf16 WER regression on that slice.
MOSS-Transcribe-preview-2B-all-4bit-g64/	4-bit affine, group 64 (all weights)	1.3 GB	Full-model 4-bit; retained as a size-optimal candidate.

Each variant ships config.json, the Qwen ByteLevel tokenizer files, and the conversion/quantization reports produced by the converter.

Validation (Apple Silicon, conversion project)

• LibriSpeech clean-test 100-row baseline: 1.80% WER, 1.61 RTFx (MLX bf16), vs 2.01% WER on the PyTorch bf16 reference.
• Weight-load, fixture-transcription, and streamed-eval gates passed on Apple Silicon.

Usage

These weights are not loadable by vanilla mlx-lm. MOSS-Transcribe is a multimodal audio-encoder + adapter + Qwen3-decoder architecture with no upstream mlx-lm model type. Load them with the conversion project's runtime:

git clone https://github.com/chikingsley/peacock-asr
cd peacock-asr/projects/moss-mlx-conversion
uv run moss-streaming-eval --weights <variant-dir>   # streamed LibriSpeech WER eval

See the project README.md and docs/PROGRESS.md for the smoke test (moss-mlx-smoke), the local backend (MossTranscribeBackend.generate), and eval harness details.

License

Apache-2.0, matching the upstream model. If you use these weights, also cite the upstream model:

@misc{moss_transcribe_2025,
  title        = {{MOSS-Transcribe-preview-2B}},
  author       = {{OpenMOSS Team}},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/OpenMOSS-Team/MOSS-Transcribe-preview-2B}}
}