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

MOSS-Transcribe-preview-2B — CoreML bundle (experimental)

CoreML 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).

Produced by the peacock-asr moss-mlx-conversion project as compiled .mlmodelc components, packaged as a FluidAudio-style model bundle.

Experimental — read before use. This is a correctness-validated research bundle, not a production FluidAudio backend. It is single-window (30-second padded static audio input), decode is autoregressive with explicit KV-cache tensors moved through CoreML every token, and it is far slower than FluidAudio Parakeet-class ASR. Only 20 LibriSpeech clean-test rows were validated through the FluidAudio-shaped runtime. No production parity is claimed.

Contents

Path	Component	Size
compiled/moss_token_embedding.mlmodelc	Token embedding	594 MB
compiled_audio_30s/moss_audio_encoder_adapter_30s_padded.mlmodelc	Audio encoder + adapter, padded 30 s mel input [128, 3000] with masked invalid audio tokens	1.4 GB
compiled_prefill_cache_512/moss_decoder_prefill_cache_512.mlmodelc	Full decoder prefill, 512-token padded cache	3.3 GB
compiled_step_padded/moss_decoder_step_padded_fixture.mlmodelc	Padded external-cache decoder step (768-slot cache)	3.3 GB
compiled_step_padded_512/moss_decoder_step_padded_512.mlmodelc	Padded external-cache decoder step (512-slot cache)	3.3 GB
compiled_prefill_cache_768/	Matched 768-token prefill — experimental, crashes in CoreML/MPSGraph cpu-gpu execution	3.3 GB
moss_bundle_manifest.json	Bundle metadata: package paths, tokenizer/runtime paths, cache presets	—
moss_runtime_manifest.json	Prompt template and model constants	—
moss_tokenizer.json	Qwen ByteLevel tokenizer	—

Cache presets (from moss_bundle_manifest.json)

• short-512: 512-prefill + 512-cache step. Validated on the 20-row short gate.
• compat-768 (default): 512-prefill + 768-cache step. Validated on the 20-row compat gate.
• matched-768: experimental; Torch-validates but fails in CoreML/MPSGraph cpu-gpu execution. Do not use.

Validation

• 20-row LibriSpeech clean-test gate through the FluidAudio-shaped Swift runtime: WER 0.0158, CER 0.00418, 0.69–0.77 RTFx (single row: WER/CER 0.0 on 6930-75918-0001).
• All exported packages were validated against PyTorch fixture tensors and compiled with xcrun coremlcompiler.
• The padded 30 s audio package matches fixture generated IDs exactly under --compute-units cpu-gpu; default .all dispatch fails on ANE for the audio package — use cpu-gpu.

Known limitations

• Single-window only: capped by the 30-second static audio package; no long-audio chunking/stitching.
• English-only ASR (upstream limitation).
• Prompt length + decode budget must fit the selected cache bucket or the run fails.
• ~0.7 RTFx at best — usable for evaluation and teacher/reference work, not for production transcription.

Usage

These compiled packages are consumed by a Swift/CoreML runtime (see the conversion project's swift/ runner and the private FluidAudio scaffold patch). A runner must load the .mlmodelc bundles with MLModel, drive prefill and the padded decoder step with explicit KV-cache arrays, and apply the prompt template from moss_runtime_manifest.json:

fluidaudiocli moss-benchmark --model-dir <this-bundle> --cache-preset short-512 ...

License

Apache-2.0, matching the upstream model. Please cite the upstream model (OpenMOSS Team, 2025) when using these weights.