Google released DiffusionGemma, an experimental open model that achieves up to 4x faster inference on dedicated GPUs through parallel text generation. The 26B Mixture of Experts model breaks from traditional autoregressive approaches by generating entire text blocks simultaneously rather than token-by-token processing. Released under Apache 2.0, it targets researchers exploring speed-critical applications.

Unlike sequential LLMs that underutilize hardware during local inference, DiffusionGemma drafts 256 tokens in parallel with bi-directional attention. Performance benchmarks show 1000+ tokens per second on NVIDIA H100 and 700+ tokens per second on RTX 5090, while fitting within 18GB VRAM through quantization. The model activates only 3.8B parameters during inference, making it accessible for consumer hardware.

However, this speed comes with trade-offs. Output quality falls below standard Gemma 4 models, making them better choices for production applications requiring maximum fidelity. DiffusionGemma excels in non-linear domains like in-line editing, code infilling, and iterative refinement tasks where bidirectional context matters more than absolute quality.

The model supports fine-tuning for specialized applications, with examples showing Sudoku solving capabilities that traditional autoregressive models struggle with. Developers can integrate via Hugging Face Transformers, vLLM, MLX, and upcoming llama.cpp support, with NVIDIA optimizations across consumer and enterprise hardware stacks.