Google unveils DiffusionGemma, delivering up to 4x faster inference on dedicated GPUs

Google has released DiffusionGemma, a new experimental open model employing text diffusion to accelerate text generation tasks. Departing from traditional sequential token-by-token methods in large language models, DiffusionGemma generates entire blocks of text at once. This leads to a significant boost in speed, with the model delivering up to four times faster output on dedicated GPUs, reaching 1,000 tokens per second on an NVIDIA H100 and 700 on a GeForce RTX 5090.

Building on the Gemma 4 architecture and Gemini Diffusion research, DiffusionGemma introduces a novel diffusion head to maximize text generation speed. The 26 billion parameter Mixture of Experts (MoE) design activates only 3.8 billion parameters during inference, making it feasible to run on high-end consumer GPUs with just 18 GB of video memory when quantized. The model supports bidirectional attention by generating 256 tokens in parallel, allowing every token to interact with all others, which is especially useful for domains like in-line text editing, code infilling, and structured data generation.

Alongside its performance improvements, DiffusionGemma refines output in real time by evaluating entire text blocks for self-correction. While suitable for researchers and developers working with speed-critical, interactive local workflows, the model remains experimental and offers lower output quality than the standard Gemma 4, which is recommended for production environments that require maximum text quality.