NVIDIA Accelerates Google DeepMind’s DiffusionGemma for Local AI

Today, Google DeepMindĀ releasedĀ DiffusionGemma — an experimental openĀ model built for exceptionally fast textĀ generation.Ā NVIDIA hasĀ optimizedĀ DiffusionGemma to run even faster across NVIDIA GeForce RTXĀ GPUs,Ā theĀ NVIDIA RTX PROĀ platformĀ and NVIDIA DGX SparkĀ systems, from local PCs to the cloud.Ā 

Rather thanĀ generatingĀ text one word at a time, DiffusionGemma generatesĀ multiple words inĀ parallelĀ to outputĀ whole blocks of text, opening a new,Ā low-latency frontier for theĀ kind ofĀ single-user workloads that developers, researchers and AI enthusiasts run every day.Ā 

Features of the new model include:Ā 

  • Parallel generation:Ā DiffusionGemma denoises upĀ to 256 tokensĀ per step instead ofĀ predicting one at a time.Ā 
  • Built on GemmaĀ 4:Ā DiffusionGemma is built on Gemma 4, aĀ 26-billion-parameter mixture-of-experts model that activates just 3.8 billion parameters per step, pairing a diffusion head withĀ Google’sĀ Gemma 4 architecture.Ā 
  • Up to 4x fasterĀ performance: The boost means fast text generation, where single-user generation usually stalls — on local hardware.Ā 
  • Open and local:Ā DiffusionGemma is openĀ weightsĀ under a permissive Apache 2.0 license and runs entirely on RTX and DGX Spark — no cloud, no per-token cost — with day-zero support inĀ Hugging Face Transformers, vLLM and Unsloth.Ā 

A Different Way to Generate TextĀ 

Almost everyĀ largeĀ language modelĀ (LLM)Ā in wide use today is autoregressive — meaningĀ it generates text one token at a time, with eachĀ new wordĀ depending on the one before it. That sequential process is what makes interactive AI feel likeĀ it’sĀ typing.Ā 

DiffusionGemmaĀ takes a different path. Built on the Gemma 4 26B mixture-of-expertsĀ architecture,Ā it generates text the way diffusion models generate images: by starting from noise and refining a whole blockĀ of textĀ at once. Each step denoises up to 256 tokens in parallel rather than emitting a single token and waiting to compute the next.Ā 

The result is a model that thinks in blocks instead of sequentially. For latency-sensitive, single-user work — such as interactive chat, agentic loops or on-device assistants that plan and act — that parallelism translates into responses fast enough to keep pace with how developers think and iterate.

DiffusionGemma Flies on NVIDIAĀ GPUsĀ 

Generating one token at a time is fundamentally a memory-bound problem — a traditional LLM spends most of its time waiting on memory bandwidth, not doing math, whichĀ leavesĀ a lot ofĀ computeĀ on the table.Ā 

Diffusion flips the equation. Pulling a full 256-token block through the transformer in parallel is a compute-bound workload — exactly what NVIDIA GPUs are built for. NVIDIA Tensor Cores accelerate the dense parallel math, and the CUDA software stack lets the model run efficiently from day one without bespoke tuning.Ā In short, theĀ model’sĀ design plays directly to theĀ GPUā€™ā€˜s strengths.Ā 

That shows up in the numbers. DiffusionGemma delivers 1,000 tokens/sec on a single NVIDIA H100 Tensor Core GPU, 150 tokens/sec on NVIDIA DGX Spark and up to 2,000 tokens/sec on NVIDIA DGX Station — roughly 4x faster than an equivalent autoregressive model running in the same single-user regime.

That advantage holds across NVIDIA’s full lineup, running:Ā 

  • Locally onĀ theĀ NVIDIAĀ DGX SparkĀ desksideĀ personalĀ AIĀ supercomputer — powered by theĀ NVIDIAĀ GB10 Grace Blackwell Superchip with 128GB of unified memory — with the preinstalled NVIDIA AI software stack ready for prototyping, fine-tuning and fully local agent workflows.Ā 
  • On NVIDIA RTX PRO 6000 workstations,Ā providingĀ developers, researchers and AI professionalsĀ withĀ theĀ headroom to run local low-latency generation and agentic loops as part of a professional workflow.Ā 
  • On DGX Station, delivering best-in-class, local high-speed inference with up to 2,000 tokens/sec for low-latency text generation and agentic loops with 748GB of coherent memory.
  • On GeForce RTX GPUs, withĀ llama.cpp support coming soon.Ā 

Get Started Locally

The fastest way to start testing and prototypingĀ the model is throughĀ Hugging Face Transformers, which runs DiffusionGemma on a GeForce RTX 5090 or DGX Spark out of the box. For higher-throughput inference, vLLMĀ provides day-zeroĀ serving support.Ā Ā 

For adapting the model to a specific task or domain, fine-tuning is available throughĀ UnslothĀ and NVIDIAĀ NeMoĀ framework, with ready-made DGX Spark playbooks to get a local environment running quickly.Ā Check out theĀ vLLMĀ playbooks forĀ DGX SparkĀ ,Ā RTX PROĀ andĀ DGX Station.Ā 

Try Diffusion Gemma on Hugging FaceĀ or test it for free using NVIDIA-hostedĀ application programming interfacesĀ atĀ build.nvidia.com.Ā 

Go deeper on the architecture and local deploymentĀ byĀ readingĀ theĀ NVIDIA technical blog and the Google DeepMind announcement.

#ICYMI: The Latest From RTX AI GarageĀ 

šŸŽ¬Ā NVIDIA researchers released SANA-WM, an openĀ source world model that turns a single image and a camera path into a minute-long, 720p video with precise 6-DoF control. At just 2.6 billion parameters, its distilled version generates a full 60-second clip in 34 seconds on a single NVIDIA GeForce RTX 5090Ā GPUĀ usingĀ theĀ NVFP4Ā format — delivering up to 36x higher throughput than comparable open models while running on one GPU. ReadĀ the paper.Ā 

šŸ› Ā Building Windows agents just got a full toolset — NVIDIA and MicrosoftĀ rolled out turnkey agent sandboxing on native Windows — MicrosoftĀ eXecutionĀ Containers plus the NVIDIAĀ OpenShellĀ runtime — alongside up to 2x faster agentic inference and native Windows support for Hermes Agent.Ā 

šŸ¤–DGX Spark goes from unboxing to a running agent in minutes — A streamlined NVIDIAĀ NemoClawĀ install gets developers to a working local agent fast, with Qwen3.6-35B running up to 2.6x faster onĀ vLLM. And the new cluster assistant in NVIDIA Sync links up to four DGX Spark units into one 512GB pool — enough for ~400-billion-parameter models.Ā 

Plug in to RTX Spark onĀ Facebook,Ā Instagram,Ā TikTokĀ andĀ X — and stay informed by subscribing to theĀ RTX Spark newsletter.Ā 

SeeĀ noticeĀ regarding software product information.

Cart

Your Cart is Empty

Back To Shop
Continue Shopping
Payment Details
Sub Total 0,00 €