NVIDIA's Two-Tower diffusion LLM generates text 2.4x faster

NVIDIA's Two-Tower diffusion language model generates text in parallel blocks, achieving 2.4x speed gains with 98.7% quality retention compared to token-by-token generation. The architecture uses a two-tower design for efficient parallel decoding.