DiffusionGemma: Google & NVIDIA's Fast AI Model Released

Google has officially released DiffusionGemma, an experimental open-source language model. What's interesting is this model introduces the diffusion mechanism, previously used in image AI, into text generation. Through optimizations by NVIDIA, DiffusionGemma runs nearly four times faster than similar traditional models in a single GPU single-user mode. It can output 1000 tokens per second on an H100 graphics card, and over 700 tokens per second on an RTX5090. This model features a "full-block awareness" capability, allowing all tokens to refer to each other during generation. This makes it particularly strong in tasks like text completion, code filling, and Sudoku solving. Its weights are open-sourced on Hugging Face under the Apache 2.0 license. This development could open new possibilities for AI in complex logic and nonlinear text generation.