Google open model · checked July 30, 2026 Experimental · Apache 2.0

Quick answer

DiffusionGemma trades quality for parallel local speed.

Google's experimental 26B-total MoE refines 256-token blocks instead of decoding every token in sequence. Test it for interactive local workflows; keep standard Gemma 4 as the quality baseline.

Verified model record

Model status

Experimental open-weights model. Google recommends standard Gemma 4 when maximum output quality matters.

Architecture

26B-total sparse Mixture of Experts built on Gemma 4 and Gemini Diffusion; Google describes 3.8B active parameters and rounds that to 4B in developer documentation.

Generation method

Discrete text diffusion that iteratively refines 256-token blocks instead of decoding every output token strictly left to right.

Input and output

Accepts text, image, and video inputs and produces text. Google's developer overview says audio input is not supported.

License and weights

Apache 2.0 open weights distributed through Google's Hugging Face and Kaggle collections, with Model Garden access also listed.

Flowith access

Not asserted here. Open weights and Google deployment surfaces do not establish availability in the live Flowith workspace.

Workload map

Use the generation method where it changes the task

Interactive local editing

Promising test case

Parallel, bidirectional refinement can help inline editing, code infilling, structured blocks, and other tasks where future tokens matter.

Low-concurrency dedicated GPU

Where speed claims are most relevant

Google says the throughput advantage is strongest at low-to-medium batch sizes on one accelerator, not as a universal cloud-serving advantage.

Maximum production quality

Prefer standard Gemma 4

Google explicitly says DiffusionGemma's overall output quality is lower than standard Gemma 4 because it prioritizes speed and parallel layout generation.

High-QPS cloud serving

Benchmark carefully

Autoregressive models can saturate cloud compute through batching; Google says diffusion's advantage can diminish and serving cost can rise.

Source-safe performance read

Google reports more than 1,000 tokens per second on one NVIDIA H100 and more than 700 on an RTX 5090 in its test configuration, with up to 4x faster output. Those measurements depend on dedicated hardware, quantization, implementation, sampling settings, prompt shape, and output length.

Record time to first usable block, completed-task latency, output quality, correction loops, memory use, power, and accepted-result cost. Raw tokens per second alone cannot choose a production model.

Official sources

  1. Google DeepMind model page — model role, architecture, speed, hardware, and downloads.
  2. Google launch announcement — license, measured speed, quality, concurrency, and runtime boundaries.
  3. Google AI developer overview — modalities, sampling configuration, and deployment links.

Continue by decision intent

Local hardware and quantization guide Turn the 18GB quantized target into a measured deployment plan. Diffusion vs autoregressive guide Choose by workload, quality target, concurrency, and accepted-result cost. Gemma 4 family Compare the standard open-model family Google recommends for higher-quality output. Gemma 4 12B Review a different local multimodal target without inheriting DiffusionGemma claims.

DiffusionGemma questions, answered

DiffusionGemma is Google's experimental open-weights text-diffusion model. It is built on Gemma 4 and Gemini Diffusion research and generates text by iteratively refining token blocks in parallel.
Not universally. Google reports strong speed on dedicated GPUs at low-to-medium concurrency, but also says overall output quality is lower than standard Gemma 4 and the advantage diminishes in highly batched cloud serving.
Google says a quantized configuration can fit within an 18GB VRAM target. Actual usage depends on checkpoint format, quantization, runtime, prompt and output length, caches, multimodal inputs, and system overhead.
Yes. Google's developer overview says it accepts text, image, and video inputs and generates text output. Audio input is not supported.
Google released the weights under Apache 2.0. Review the exact repository, notices, dependencies, acceptable-use requirements, and deployment context before shipping.
This page does not claim a Flowith model mapping. Verify the live Flowith workspace selector separately from Google's open weights, Model Garden, and third-party runtime support.