Q8 quantized MLX conversion of whisper-large-v3-swiss-german, derived from
openai/whisper-large-v3 plus a Swiss German PEFT adaptation trained on
SwissDial data.
This repository is intended for local speech-to-text inference with MLX-based
Whisper runtimes, including oMLX through its mlx-audio STT integration.
Format
Native MLX Whisper checkpoint
Weight-only affine quantization
8-bit weights
Group size 64
model.safetensors with MLX tensor names
Hugging Face Whisper processor/tokenizer files included for mlx-audio/oMLX
Usage
Download the repository and point your MLX/oMLX Whisper STT engine at the local
model directory or Hugging Face repo:
For oMLX, configure the model as a Whisper/STT model. Exact routing depends on
the oMLX version and deployment config.
License and usage restrictions
SwissDial CC BY-NC 4.0 with no-reidentification condition
The Swiss German adaptation is based on SwissDial data. Use of this model is
therefore restricted to non-commercial purposes under Creative Commons
Attribution-NonCommercial 4.0 International, plus the SwissDial condition that
users must not attempt to determine the identity of speakers in the SwissDial
dataset.
Additional practical terms:
Non-commercial use only.
Attribute SwissDial and cite the SwissDial publication when used in research.
Do not use this model or its outputs to identify or re-identify SwissDial
speakers.
This model is not endorsed by ETH Zurich, OpenAI, or the original adapter
authors.
The Whisper base model remains under its original Apache-2.0 license. The
combined adapted checkpoint is distributed under the more restrictive terms
above because of the SwissDial-trained adaptation.
Provenance
This Q8 model was produced from the local FP16 MLX checkpoint by quantizing
MLX Linear and Embedding weights to 8-bit affine groups. Conv, LayerNorm,
positional embeddings and bias vectors remain dense.
See merge_provenance.json for the base and adapter source metadata inherited
from the FP16 conversion.
Validation
The repository contains SHA256SUMS for model.safetensors and config.json.
Structural validation checks:
Whisper large-v3 MLX dimensions
required processor/tokenizer files
required representative encoder/decoder tensors
quantization metadata in config.json
No benchmark numbers are claimed for this Q8 variant until dialect-specific
WER/CER evaluation has been run on a held-out Swiss German test set.
