DiffusionGemma 26B-A4B

DiffusionGemma 26B-A4B, released June 10, 2026 by Google DeepMind, is an experimental open-weights model that replaces token-by-token autoregressive decoding with discrete diffusion: it generates 256-token canvases in parallel through iterative denoising. It shares the Gemma 4 26B A4B MoE backbone (25.2B total, 3.8B active) and delivers up to 4x faster throughput for single-user, low-batch workloads, with native vLLM support from day zero. A community site at diffusiongemma.dev tracks benchmarks.

The honest picture: this is a research artifact, not a production model. DiffusionGemma scores lower than the autoregressive Gemma 4 baseline across reasoning, coding, and multimodal evals — the speed is a tradeoff, not a free upgrade. The speed advantage also collapses on low-compute GPUs and Apple Silicon. Its real value is giving the open field its first serious test case for diffusion-based language modeling.