DiffusionGemma: 4x faster text generation

Google Blog (General)

Google's DiffusionGemma generates 256 tokens simultaneously using diffusion, achieving up to 4x faster inference than autoregressive models. The 26B MoE model (3.8B active) is open-source under Apache 2.0, targeting speed-critical local AI workflows. It runs on consumer GPUs with 18GB VRAM when quantized, delivering 1000+ tokens/s on an H100.

GemmaDiffusionGoogle

See 3 more sources

What Is Diffusion Gemma? Google's Text Model That Generates 256 Tokens at Once3 days agoMindStudio Team

What Is Google Diffusion Gemma? The Text Model That Generates 256 Tokens at Once3 days agoMindStudio Team

Diffusion Language Models Explained: How Google's Diffusion Gemma Works3 days agoMindStudio Team

Google Blog (General)·Sebastian Flennerhag

8 days ago