Google DeepMind Releases DiffusionGemma for High-Speed Text Generation

Google DeepMind released DiffusionGemma, an experimental open-weight AI model built on the Gemma 4 architecture. Unlike traditional autoregressive models that generate text sequentially, DiffusionGemma uses a diffusion-based approach to produce and refine blocks of up to 256 tokens simultaneously. This parallel processing allows for inference speeds up to four times faster than comparable models, particularly in local, single-user scenarios.

The 26-billion-parameter Mixture-of-Experts model activates 3.8 billion parameters per step and is released under an Apache 2.0 license. It is specifically designed for speed-critical workflows such as inline editing, agentic loops, and interactive coding. While offering significant latency advantages and the ability to fit within 18GB of VRAM on high-end consumer GPUs, Google noted that the model's overall output quality is lower than standard Gemma 4 models and provides diminishing returns in high-QPS cloud environments.

NVIDIA provided day-one support and optimizations for the model to leverage GPU Tensor Cores and CUDA, shifting workloads from memory-bandwidth bottlenecks to compute-intensive processing. Performance metrics show a single H100 GPU on a DGX Station can reach 1,000 tokens per second. The model is available via Hugging Face, GitHub, Google Cloud Model Garden, and Nvidia NIM, with additional fine-tuning support from Unsloth for specialized tasks like solving Sudoku puzzles.