Google has released an experimental open-source text model called DiffusionGemma that ditches the usual word-by-word generation in favor of a diffusion-based approach — the same technique behind AI image generators like Stable Diffusion. Instead of predicting one token at a time, DiffusionGemma starts with random placeholder tokens and iteratively denoises them over multiple steps, producing entire blocks of text in parallel. The result? Up to 4x faster inference on a single consumer GPU.
DiffusionGemma is a 26-billion-parameter model, but thanks to a Mixture-of-Experts (MoE) architecture, it only activates about 3.8 billion parameters per forward pass. During each pass, the model generates 256 tokens simultaneously, enabling bidirectional attention across the entire block. This makes tasks like code completion, inline editing, and math formula generation much more natural — though Google admits the overall output quality still trails its standard Gemma 4 model.
On the hardware front, Google says a single NVIDIA H100 can push over 1,000 tokens per second, while a consumer-grade RTX 5090 manages more than 700. With 4-bit floating point (NVFP4) quantization, memory usage drops below 18GB, making local deployment much more accessible. The weights are already up on Hugging Face, and support is available for MLX, vLLM, Unsloth, and NVIDIA NeMo.