Google DeepMind released DiffusionGemma this week, an experimental open model that does something most language models do not: it generates text the way an image model generates a picture. Where a normal autoregressive LLM writes one token at a time, left to right, like a typewriter, DiffusionGemma starts from a block of random placeholder tokens and denoises the whole block at once, up to 256 tokens per forward pass, making several refining passes until the text converges. Because the whole block is present from the start, attention is bi-directional, every token can attend to every other, and the model can revise an earlier token after seeing later ones, a kind of running self-correction the left-to-right paradigm cannot do. It is a 26-billion-parameter Mixture of Experts that activates 3.8 billion per step, built on the Gemma 4 family with diffusion research carried over from Gemini Diffusion, released under Apache 2.0 with weights on Hugging Face and a quantized footprint that fits in 18GB.
The speed is the headline number and it is real: about 4x faster than an equivalent autoregressive model in single-user generation, which DeepMind clocks at more than 1,000 tokens per second on a single NVIDIA H100 and over 700 on a consumer RTX 5090. The honest caveat is stated by Google itself, not buried: DiffusionGemma's overall output quality is lower than standard Gemma 4, and for quality-critical work the recommendation is to use standard Gemma 4 instead. So this is not a free upgrade. It is a speed-for-quality trade, shipped as an experiment you can actually run, and the right way to read it is as a research artifact with usable weights rather than a Gemma 4 replacement.
The part worth a builder's full attention is why it is fast, because it inverts the usual rule. Autoregressive decoding is bottlenecked on memory bandwidth: each token requires reading the whole model, so you are limited by how fast weights move, not how fast the chip computes. Generating a 256-token block in parallel flips that, the bottleneck moves from memory bandwidth to compute, exploiting the high arithmetic intensity that modern accelerators are starved to use. The practical consequence is counterintuitive: DiffusionGemma wins on local, low-concurrency hardware, a single user on one GPU, and is not built for high-QPS cloud serving where batching already keeps the accelerators busy. DeepMind even notes that memory-bandwidth-bound machines like Apple Silicon Macs may not see the same gain. It runs across Hugging Face Transformers, vLLM, MLX, and NVIDIA NeMo, with llama.cpp support coming.
For the Gemma velocity arc we have been tracking, this adds a new axis. The story so far was the same model getting smaller and faster-quantized, laptop-local then phone-local. DiffusionGemma is a different move: not a smaller Gemma, a differently-generated one, the first time the diffusion paradigm that powers image and video generation ships as a usable open text model you can fine-tune and run. Its natural home is the set of tasks where tokens depend on tokens that come later, code infilling, structured or non-linear text, even the toy case DeepMind shows of a fine-tune solving Sudoku, where left-to-right is exactly the wrong shape. It is experimental and lower-quality today. But a generation paradigm that flips the decode bottleneck from memory to compute is the kind of thing that matters more as local inference becomes the place builders actually want to run, and that is the lane this is aimed straight down.