NVIDIA NemotronLabs RE-USE trends on Hugging Face with 124 likes
NVIDIA's NemotronLabs RE-USE model appeared among the most-liked models on Hugging Face today with 124 likes and 29,275 downloads.
Model details
The model is a 9.6-million-parameter multilingual universal speech enhancement model, according to NVIDIA. Its architecture includes a convolutional encoder, convolutional decoder, and a bi-directional Mamba with 30 layers for time–frequency modeling, according to NVIDIA. It supports input sampling rates of 8, 16, 22.05, 24, 32, 44.1, and 48 kHz, according to NVIDIA. Input audio must be .wav files, mono-channel, 1D, sampled between 8000 Hz and 48000 Hz, and output audio follows the same format, according to NVIDIA.
Deployment and hardware
The model is ready for commercial use and was released on 10/22/2026, according to NVIDIA. Supported hardware microarchitecture is NVIDIA Ampere and the supported operating system is Linux, according to NVIDIA. Inference uses the Triton acceleration engine on NVIDIA Ampere A100 GPUs, according to NVIDIA. The cheapest GPU rental in our data that fits the model is a Quadro P2000 with 5 GB VRAM at $0.0289/hr from vast.ai.
Benchmark claim
On the full test set, NemotronLabs RE-USE achieved a DNSMOS score of 3.29, according to NVIDIA.
Key takeaways
- NVIDIA's compact speech enhancement model is trending on Hugging Face.
- It runs on Ampere GPUs with Triton acceleration and is cleared for commercial use.
- NVIDIA reports a DNSMOS score of 3.29 on the full test set.
Official announcement
Published by nvidia. Read it at the source.