DiffusionGemma

Google has released the diffusion version of Gemma 26BA4B. The model is 4 times faster with the same number of parameters — on H100 it outputs 1000 tokens per second, and on RTX 5090 a very decent 700. The quality decreases compared to the autoregressive model, but training text diffusion is still a learning process, which is not surprising. It's nice to see that work on diffusion for LLMs continues, as the demo of Gemini Diffusion a year ago never made it to release.