← All models

DiffusionGemma

Published Publisher Google

DiffusionGemma

Positioning and release

  • Released: June 10, 2026.
  • Positioning: An experimental, open-weights model from Google DeepMind for research and speed-sensitive interactive workflows, built on Gemma 4 and Gemini Diffusion research. The weights are released under the Apache 2.0 license.
  • Release identifier: diffusiongemma-26B-A4B-it. This is the supplied release identifier, not a verified Google API model ID. (blog.google)

Architecture and capabilities

  • A mixture-of-experts model with 25.2 billion total parameters and 3.8 billion active parameters. It uses block-autoregressive discrete diffusion to iteratively refine 256-token blocks, with up to 256K tokens of context.
  • Accepts text, images and video frames and generates text. Audio input is not supported. It supports configurable thinking, coding, structured function calling and multilingual tasks.
  • Google reports up to 4× faster token generation on dedicated GPUs, with more than 1,000 tokens per second on a single NVIDIA H100 and more than 700 on an RTX 5090. These are reported results, not guaranteed speeds; performance depends on the workload and hardware. Quantized deployment can fit within 18 GB of GPU memory. (ai.google.dev)

Uses and deployment

Suitable for experiments involving local interactive assistants, inline editing, code completion, rapid text iteration, image and video understanding, and non-linear text generation. Google documents local serving with vLLM and deployment through Google Cloud Model Garden. (developers.googleblog.com)

Limitations

  • Google describes the model as experimental and says its overall output quality is lower than standard Gemma 4; it recommends standard Gemma 4 when maximum production quality is required.
  • Speed advantages are strongest at low to medium batch sizes. They may diminish at high cloud concurrency or on hardware with different compute and memory characteristics.
  • Outputs can be incorrect, outdated or biased. Video is processed as frames; the model card specifies a maximum of 60 seconds at one frame per second. (blog.google)

Pricing

Unverified: As of November 16, 2026, a Google-published per-token API price for this specific model could not be verified. Input price, output price and billing currency are therefore unverified. The weights are open, but self-hosting or cloud deployment can incur infrastructure charges; no amount is asserted here. The official Gemini Developer API pricing page was checked and did not list DiffusionGemma. (aistudio.google.com)

Official sources