NVIDIA Releases Open-Source Dual-Tower AI Model, Text Generation Speed Increased by 2.42 Times, Image Quality Retained at 98.7%

NVIDIA released the Nemotron-Labs-TwoTower discrete diffusion language model, solving the problem of slow token-by-token generation speed in large models. The weights have been open-sourced on Huggingface. The model reuses pre-trained weights of existing backbone networks without the need for retraining from scratch, significantly reducing costs. It adopts a 60B dual-tower architecture, with two 30B networks working in parallel. Each tower activates 3B parameters and is equipped with 128 routable expert modules to improve generation efficiency.