What is different
Google released DiffusionGemma on June 10, 2026. Instead of generating every token strictly from left to right, it starts with a block of draft tokens and refines multiple positions across diffusion steps. The model has 26B total MoE parameters, activates 3.8B per token, and has a quantized configuration intended to fit in about 18 GB of VRAM.
Is it free to use?
Google published downloadable weights on Hugging Face under Apache 2.0. There is no per-token fee for the weights, but users still carry repository access requirements, memory, GPU electricity, and hosting costs.
Reported advantages
- Parallel token refinement targets lower latency at low or medium batch sizes on a single accelerator.
- Google reports generation up to four times faster on GPUs.
- The process can support research into editing intermediate generations and exploring candidate outputs in parallel.
Important limitations
Google explicitly says overall output quality is lower than standard Gemma 4 and recommends Gemma 4 when maximum production quality matters. At high query volume and large batches, the speed advantage may shrink or cost more. Builders should compare first-token latency, full response time, quality, and throughput per watt on identical hardware.