DiffusionGemma Technical Report

DiffusionGemma generates text by repeatedly refining a block of token predictions. An ordinary autoregressive model commits one token before predicting the next. Here, positions inside a block can influence one another during denoising, allowing many output tokens to be resolved with fewer model calls.

The starting point is Gemma 4's mixture-of-experts model. Supervised adaptation teaches block denoising; subsequent sampler distillation and reinforcement learning improve quality while reducing the required refinement steps. The result retains autoregressive generation as another operating mode, but the diffusion mode represents a changed model distribution, not an exact acceleration of the original model.

The report's speed measurements are impressive under their stated conditions. The quality tables also show clear losses relative to the autoregressive initialization. Both observations matter. A fast response that misses a required action is not an equivalent replacement for a slower successful response.

This July 2026 paper is particularly useful for understanding why serving systems should report decoding throughput, complete-request latency and task success separately. Its maximum denoising budget, adaptive stopping behavior and workload-dependent output lengths make a single tokens-per-second number an incomplete deployment specification.

An interactive assistant may spend much of its time moving weights and attention state rather than doing useful arithmetic. Generating several positions together can use that hardware more effectively. That makes diffusion worth investigating for latency-sensitive generation, especially when the output contains predictable structure or copies from the input.

The customer still experiences queuing, prompt processing, tool calls and validation. A system that finishes text quickly can remain slow to produce its first useful answer. It can also require more correction if the model's instruction following or tool-use success changes.

Start with inference serving and continuous batching. Compare this paper with speculative decoding: a verified draft can preserve a target distribution, while direct diffusion generation trades into a different quality-speed operating point. That distinction determines what must be revalidated before changing a customer's model.