Tag
The paper proposes Nemotron-TwoTower, a diffusion language model that decouples context representation and denoising using a frozen autoregressive tower and a trainable diffusion denoiser, achieving 98.7% of baseline quality with 2.42x throughput.