NVIDIA's compressed Puzzle-75B and audio-text Audex Nemotron models climb Hugging Face trending list

Two NVIDIA models published quietly to Hugging Face in early July have climbed onto the platform’s trending list this week. Leading the surge is Nemotron-Labs-3-Puzzle-75B-A9B, a compressed model NVIDIA says runs faster than the larger model it was built from.

The two model cards first appeared between July 6 and July 9 and reached the trending models list at #20 for Puzzle-75B and #17 for the companion audio-text model by July 11 — a slow-burn debut for releases that arrived without a launch push.

Puzzle-75B is a hybrid mixture-of-experts (MoE) model derived from Nemotron-3-Super-120B-A12B using NVIDIA’s Iterative Puzzle compression framework, which cuts the model from 120.7 billion total parameters to 75.3 billion, and from 12.8 billion to 9.3 billion active. NVIDIA claims the compressed model delivers up to 2.03x the server throughput of its parent on an eight-GPU B200 node at matched user throughput.

Those figures are NVIDIA’s own and have not been independently verified. Throughput gains at “matched user throughput” hinge on batch sizes and serving configuration that real-world deployments may not replicate.

The companion release, Nemotron-Labs-Audex-30B-A3B, is a unified audio-text MoE model with 30 billion total and 3 billion active parameters and up to a 1-million-token context window, trained on 157.4 billion audio tokens and 320.5 billion text tokens. Audex ships under NVIDIA’s noncommercial license, which limits use to research and evaluation, while Puzzle-75B carries the more permissive OpenMDW-1.1 license.