NVIDIA's Two-Tower diffusion LLM achieves 2.4x speed gains

NVIDIA's Two-Tower diffusion LLM generates text in parallel blocks instead of token-by-token, achieving 2.4x speed gains. The model retains 98.7% of the quality of autoregressive LLMs.