Version 1

Google DeepMind released DiffusionGemma on June 10, 2026, a 26B mixture-of-experts text model that generates up to 256 tokens simultaneously by denoising from noise rather than predicting one token at a time. [1] The model achieves roughly 4x faster output than comparable autoregressive models — over 1,000 tokens/sec on a single H100 and over 700 tokens/sec on an RTX 5090. [1] NVIDIA simultaneously published acceleration guidance, framing the model's parallel generation pattern as a natural fit for GPU compute strengths. [2] DeepMind released the model as open weights under Apache 2.0 with day-zero support in Hugging Face Transformers, vLLM, and Unsloth, while explicitly acknowledging output quality is lower than standard Gemma 4. [1]

DiffusionGemma is the most prominent open release to test whether diffusion-based text generation can deliver practically useful speed gains outside the lab. The quality trade-off DeepMind openly acknowledges makes it a research and interactive-use model for now, but if that gap narrows, the architectural approach could reshape how inference is planned for local and GPU-server deployments.

Google DeepMind released DiffusionGemma on June 10, 2026, as an experimental open-weights model built on the Gemma 4 26B mixture-of-experts architecture. Unlike autoregressive language models that predict one token at a time, DiffusionGemma generates text the way diffusion models generate images: it starts from a block of noise and iteratively refines all 256 tokens simultaneously. [1] Because the model activates only 3.8B of its 26B parameters per inference step, it fits within 18GB of VRAM when quantized, making it accessible on consumer hardware. [1]

The central performance claim is roughly 4x faster token output in single-user settings compared to equivalent autoregressive models. DeepMind reports over 1,000 tokens/sec on a single NVIDIA H100 and over 700 tokens/sec on a GeForce RTX 5090. [1] NVIDIA, which co-publicized the launch, explains the speed advantage in structural terms: pulling a full 256-token block through the transformer at once converts inference from a memory-bandwidth-bound workload into a compute-bound one, which is where GPU silicon excels. [2] NVIDIA framed DiffusionGemma's release as day-zero support across Hugging Face Transformers, vLLM, and Unsloth. [2]

DeepMind's own announcement is notably candid about limitations. The post directly states that DiffusionGemma's output quality is lower than standard Gemma 4 and that applications requiring maximum quality should use the standard model instead. [1] The intended use cases are developer experimentation, interactive local workflows, and tasks where bi-directional attention provides a structural edge — such as code infilling and in-line text editing — rather than production deployments. [1] NVIDIA's coverage, while technically detailed, is promotional in tone and does not dwell on quality trade-offs. [2]

Beyond the two primary announcements, a body of research on accelerating diffusion language model inference already exists — covering KV caching for diffusion models, the dInfer inference framework, and NVIDIA's Torch-TensorRT optimizations for diffusion workloads [3][4][5] — indicating DiffusionGemma enters an active research space, not a blank one. The model's release as open weights under Apache 2.0 means the community can begin independent benchmarking and fine-tuning immediately.