DiffusionGemma delivers 4× faster on‑device text generation via parallel diffusion
Google’s DiffusionGemma model replaces token‑by‑token decoding with diffusion‑based parallel generation, achieving up to 4× faster inference (700+ tokens/sec on RTX 5090, 1000+ tokens/sec on H100). The 26B‑parameter MoE runs within 18 GB VRAM, making real‑time, on‑device generation feasible.