Nemotron-Labs-Diffusion: NVIDIA's Tri-Mode Unified Language Model

nvidia · hf · 2026-07-08

NVIDIA introduced Nemotron-Labs-Diffusion, a language model fusing three paradigms: Autoregressive, Diffusion, and Self-Speculation Decoding. It outperforms existing models in throughput and inference efficiency. By unifying these three decoding mechanisms, the model can flexibly switch between different inference scenarios to achieve more efficient text generation.

Original post →

More from Models

Models channel →