Google DeepMind has released DiffusionGemma, a new member of the Gemma 4 open model family that breaks from the autoregressive norm. Instead of generating text left to right, one token at a time, DiffusionGemma works like an image diffusion model: it starts with a field of placeholder tokens and iteratively denoises the entire canvas, refining all tokens in parallel until a coherent block of text emerges. The result is a 4x speed improvement over similarly sized autoregressive Gemma models, producing over 1,000 tokens per second on a single Nvidia H100 and roughly 700 tokens per second on an RTX 5090.

Architecturally, DiffusionGemma is a Mixture of Experts model with 26 billion total parameters, but only 3.8 billion are activated per inference. This keeps memory requirements under 18 GB, making it deployable on consumer-grade GPUs. The shift from autoregressive to diffusion-based generation moves the performance bottleneck from memory bandwidth to compute, enabling up to 256 tokens to be generated simultaneously.

This design choice improves performance on non-linear tasks where future tokens influence earlier ones—such as inline editing, molecular sequencing, and mathematical graphing. Google’s demo shows DiffusionGemma solving Sudoku puzzles, a notoriously difficult problem for autoregressive models because each token depends on later context. The model’s ability to self-correct across the entire token set during the denoising process makes this feasible.

For practitioners, the trade-off is clear: you gain throughput and parallelization at the cost of increased compute demand and the need to tune the denoising schedule. This isn’t a drop-in replacement for every text generation pipeline, but for workloads that benefit from bidirectional context, it’s a compelling architectural shift.

Source: Ars Technica