DiffusionGemma

DiffusionGemma is Google DeepMind's first open-weight discrete diffusion language model, released June 10, 2026. Built on the Gemma 4 26B-A4B mixture-of-experts backbone (26B total, ~3.8B active per forward pass), it drops autoregressive decoding: instead of predicting one token at a time, it starts from a 256-token canvas of noise and denoises the whole block in parallel, reaching over 1,000 tokens per second on a single H100. The context window is 256K tokens, inherited from the Gemma 4 line.

Google applies its earlier Gemini Diffusion research to the open Gemma line here, and is unusually direct about the tradeoff: DiffusionGemma scores below standard Gemma 4 on benchmarks like MMLU and coding, and is recommended for latency-sensitive workloads where speed matters more than peak quality. It shipped with day-one support in Transformers, vLLM, and NVIDIA RTX/DGX platforms, all under Apache 2.0.