Google's new model writes text in whole blocks, not word by word — and 4× faster

Here's a curious one if you're poking at local models. On June 10 Google DeepMind released the open model DiffusionGemma — and it writes text unlike everything else.

What happened

A normal model types text one word at a time, left to right — the way you type a message yourself. DiffusionGemma works differently: it starts from "random noise" of placeholder tokens and, over several passes, refines a whole block at once (up to 256 tokens together). It's the same trick image generators use — except text, not pixels, emerges from the noise.

The model is open (Apache 2.0), the weights are on Hugging Face, and it runs in vLLM, Transformers, MLX and Unsloth.

Why it matters

Because the model refines a chunk in parallel instead of token by token, it's faster:

• up to 4× faster than the usual approach;
• 1000+ tokens/sec on an H100 GPU, 700+ on a home RTX 5090;
• quantized, it fits in 18 GB of VRAM — so it really runs on a good home card, no cloud.

What's in it for you

It's a free way to feel out a new approach to generation and get a fast local helper — for code infilling and quick iteration, where response speed matters.

Honestly, no spin: Google itself says it's not for maximum quality in production — it's for speed. So treat it as a quick "rough-draft" model on hand, not a replacement for the top ones.

No hype: not a "killer" of normal models, just a different tool — for when you need speed over a perfect answer.

Source: MarkTechPost, NVIDIA Blog