Google's DiffusionGemma generates 256 tokens in parallel and self-corrects as it goes

Google released DiffusionGemma, an open source diffusion-based language model built on Gemma 4 and supported in vLLM, generating 256 tokens in parallel for faster low-concurrency inference. Benchmarks show up to about 4–6x higher GPU token rates, but Google says output quality is lower than standard Gemma 4 and gains diminish in high-throughput batching.