NanoCodec: Towards High-Quality Ultra Fast Speech LLM Inference
Paper β’ 2508.05835 β’ Published β’ 2
NeMo NanoCodec 22 kHz / 1.89 kbps / 21.5 fps β MLX (safetensors)
MLX-ready safetensors conversion of NVIDIA's NeMo NanoCodec
(nvidia/nemo-nano-codec-22khz-1.89kbps-21.5fps),
the 22.05 kHz / 21.5 fps / 1.89 kbps variant. It is a neural audio codec using finite scalar
quantization (GroupFSQ, 8 groups Γ FSQ levels [8, 7, 6, 6], 32 embed dims) with a HiFi-GAN
encoder/decoder.
This repo exists so Swift-MLX / MLX consumers can load the codec weights directly β the
upstream distribution ships a .nemo archive that requires the NeMo/PyTorch stack (Linux-only in
practice). The tensors here are extracted from that archive verbatim (weight-norm kept in the
modern parametrizations.weight.{original0,original1} layout; no numerics changed). It is the
codec layer of the Gepard-1.0 streaming-TTS port
(xocialize/mlx-gepard-swift).
Contents
| file | what |
|---|---|
model.safetensors |
encoder (audio_encoder.*) + decoder (audio_decoder.*) + quantizer weights (fp32) |
codec_config.yaml |
the NeMo config the weights were exported with (rates, dims, FSQ levels) |
nvidia-open-model-license-agreement-june-2024.pdf |
the governing license (shipped per Β§3.1) |
NOTICE |
the Β§3.1 attribution notice |
Parity
Converted and validated against the NeMo reference (CPU fp32) as part of the Gepard port:
- decode (32-ch codes β waveform): max|Ξ| 6.7e-6, SI-SNR 104 dB, corr 1.000000
- encode (waveform β codes): 99.5% packed integer-exact (residual = FSQ knife-edge boundary flips + β€1 tail frame β the documented conv-edge convention).
License / Terms of Use
Governed by the NVIDIA Open Model License Agreement (v. June 14, 2024) β see the bundled
PDF and NOTICE. The model is commercially usable; redistribution of the model and derivatives
is permitted with the Β§3.1 obligations (ship a copy of the Agreement and the Notice below).
Licensed by NVIDIA Corporation under the NVIDIA Open Model License
The conversion tooling and the consuming Swift port are Apache-2.0; only these weights carry the NVIDIA Open Model License.
Citation
NeMo NanoCodec β https://arxiv.org/abs/2508.05835
Model size
0.1B params
Tensor type
F32
Β·
I32 Β·
Hardware compatibility
Log In to add your hardware
Quantized