An unofficial MLX conversion of NbAiLab/nb-whisper-large by NbAiLab (the National Library of Norway's AI lab), for running
locally on Apple Silicon with mlx-whisper. Published by the
Tibber-se organisation as a convenience for colleagues and the community. It is not affiliated with or
endorsed by NbAiLab, and comes with no support. All credit for the model goes to them.
uvx --from mlx-whisper mlx_whisper recording.m4a --model Tibber-se/nb-whisper-large-mlx-4bit --language no -f srt
Needs an Apple Silicon Mac and ffmpeg. The weights stay on your machine and so does the audio.
NbAiLab/nb-whisper-large at revision 8c6249fdeeb4dcd05e5735a4c39640607eb6e4ac (Apache-2.0).--dtype float16, -q --q-bits 4 (group size 64).weights.safetensors (named for the mlx_whisper loader on PyPI) and config.json.
Same tensors as the converter's own model.safetensors output.Word error rate on 100 clips of the FLEURS test split (Norwegian Bokmål, CC-BY-4.0), one run, on an Apple Silicon Mac:
| WER | |
|---|---|
whisper-large-v3-turbo (MLX) |
10.2% |
NbAiLab/nb-whisper-large through transformers (fp32, original weights) |
5.5% |
| This model | 5.5% (median clip 3.5%, worst clip 29%) |
Corpus WER is within about 0.3 points of the original weights. FLEURS is clean read speech, so real meetings and calls will score worse. Normalisation lowercases and strips punctuation but does not normalise numbers. A single run on 100 clips per language is a sanity check, not a benchmark.
NbAiLab/nb-whisper-large