Phonon-2 (MLX layout)
A format conversion of FermionResearch/Phonon-2, which is itself a five-value quantization-aware compression of nvidia/parakeet-tdt-0.6b-v3.
No weights were retrained here. The five-value bits are carried through verbatim;
tensors were renamed into the NeMo/MLX layout so an MLX Parakeet implementation can
read them. See ATTRIBUTION.md for the exact list of changes.
Licensing
- Weights: CC-BY-4.0 —
LICENSE-WEIGHTS-CC-BY-4.0.txt, inherited fromFermionResearch/Phonon-2. - Reader code used to unpack the original container: Apache-2.0 —
LICENSE-CODE-Apache-2.0.txt.
Attribution is required by CC-BY-4.0 and is owed to both FermionResearch (for Phonon-2) and NVIDIA (for the parakeet-tdt-0.6b-v3 checkpoint it was trained from).
What this file is
One model.safetensors holding 264 five-value packed blocks and 9 int6 tables,
with the packing described in its own safetensors metadata (phonon_packed,
phonon_int6). Everything else is plain float16.
Accuracy
FermionResearch publish their own WER figures for Phonon-2. They are not reproduced here and are not claimed by this repository. This conversion has been checked for fidelity to the reference implementation's outputs, which is a different thing from a benchmark.
- Downloads last month
- 128
8-bit