DeepMind Just Dropped "DiffusionGemma" — Text Generation via Image-Style Diffusion Model

Article automatically generated from technical news.

Another open weight model got dropped today, this one's from DeepMind, seems like a good day for the OSS geeks. Released under Apache 2.0 Instead of generating text sequentially token-by-token like almost every autoregressive model on the market, it uses a text diffusion head. - Throws a 256-token "canvas" of random placeholder noise onto the screen. - Uses Uniform State Diffusion to iteratively refine and denoise the entire block of text all at once. - Becau