Version 2

Google DeepMind released DiffusionGemma on June 10, 2026 — an open-weights 26B mixture-of-experts model that generates up to 256 tokens simultaneously by denoising from noise rather than predicting tokens sequentially. [2] The model achieves roughly 4x faster output than comparable autoregressive models: over 1,000 tokens/sec on an H100, over 700 tokens/sec on an RTX 5090, and roughly 550 tokens/sec in independent testing by Simon Willison via NVIDIA's free NIM cloud API. [2][4] DeepMind explicitly frames it as experimental and acknowledges output quality is lower than standard Gemma 4. [2] Willison traces the model to Gemini Diffusion research Google briefly previewed in May 2025 but did not publicly follow up on until this release. [4]

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 placeholder tokens and iteratively refines all 256 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. [2][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. [2] 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. [3] Independent testing by Simon Willison via NVIDIA's free NIM cloud API yielded 2,409 tokens in 4.4 seconds — roughly 547 tokens/sec — a real-world data point that is lower than peak advertised figures but hardware and load conditions were not specified. [4] NVIDIA is also hosting the model on that NIM API at no cost. [4]

DeepMind's own announcement is candid about limitations: output quality is lower than standard Gemma 4, and the model is explicitly not recommended for applications requiring maximum quality. [2] 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. [2] NVIDIA's promotional coverage does not foreground these quality limitations. [3] Ars Technica's independent coverage reports the performance figures and architectural novelty without endorsing or contesting the quality caveat. [1]

Simon Willison notes that DiffusionGemma is not a wholly new direction: it traces to Gemini Diffusion research that Google briefly previewed in May 2025 but did not publicly develop further until this release. [4] The model launches into an active research area — existing work covers KV caching for diffusion language models, dedicated inference frameworks, and GPU-level optimizations [5][6][7] — meaning community benchmarking and fine-tuning can begin immediately against a non-trivial research baseline.