From Typewriter to Printing Press: How Google Changed the Common Sense of AI Generation

On June 10, 2026, Google released a model called DiffusionGemma. In an era where we hear about 'new AI' almost every week, the reason this announcement is different lies not in the 'intelligence' of the model, but in the fact that it has changed the 'way it moves' itself. It is an experimental model that fundamentally overturns the 'generate one word at a time in sequence' method that all LLMs (Large Language Models) have adopted until now.
I have placed a summary at the beginning for those who want to read it quickly as it has become quite long.

——————————————————
Quick Read Version (Summary)
——————————————————

Google's DiffusionGemma generates text not 'one word at a time in sequence,' but '256 words all at once.' It achieves over 1,000 tokens per second on an NVIDIA H100 GPU, making it up to four times faster than conventional models of the same size.

What makes this design different is that the AI first outputs an 'entire draft' and then refines it. Conventional models typed while only being able to see the next word.DiffusionGemma can rewrite while looking at the entire paragraph.

It is currently positioned as an 'experimental model,' and its accuracy is inferior to conventional ones. However, if this design matures, the feasibility of high-speed AI assistants running on local PCs and real-time editing support tools will expand significantly.

——————————————————
The actual article starts from here.
—Approximately 2,172 characters / Takes about 5 minutes to read— ——————————————————

LLMs represented by GPT, Claude, and Gemini all operate on a design of 'predicting the next token.' When generating text, the model outputs one word (or more accurately, one token), uses that as a basis to output the next word, and then the next—repeating this task endlessly.

This process is inherently serial. The next word cannot begin until the previous one is determined. If we compare it to a typewriter, it is a state of hitting keys one character at a time and choosing the next key while looking at the previous character..

DiffusionGemma changes this.

The Story of the Printing Press

Google describes the design philosophy of DiffusionGemma as 'from typewriter to printing press.'

A printing press does not carve one character at a time, but presses the entire block of assembled characters onto paper at once. DiffusionGemma operates in a way similar to this.

It is a method where 256 tokens (roughly one or two paragraphs) are first laid out as a 'draft' all at once, and then corrected and refined multiple times.

Specifically, it works like this.

First, 256 random placeholders (noise) are lined up.
The AI takes several passes to replace them with 'correct tokens.'

It is an application of the noise removal process (diffusion) used when AI generates images to text generation.

As a result of this method, it achieves a generation speed of over 1,000 tokens per second on a single NVIDIA H100. Considering that conventional models of the same size are usually around 50 to 100 tokens per second, there is a speed difference of around 10 times.

Why is it fast?—The way hardware is used changes

One of the reasons why conventional LLMs are slow lies in how they use the GPU.
A GPU is a processor that is inherently good at 'processing large amounts of calculations in parallel.' However, models that output one word at a time in sequence barely utilize the GPU's parallel processing capability.

DiffusionGemma reverses this structure.
By processing 256 tokens simultaneously, it can utilize the GPU cores to the maximum.
Google explains this as a design that 'runs the GPU at full power.'

Another important point is the mechanism called bidirectional attention.
Conventional LLMs are unidirectional designs that 'can only refer to past tokens.' Because DiffusionGemma can refer to all 256 tokens simultaneously, it can generate while maintaining the consistency of the entire paragraph.
It can grasp the context of the text and correct it—in other words, it is capable of self-correction.

Runs on consumer GPUs

One of the reasons why this model is worth noting for people other than researchers is its hardware requirements.

DiffusionGemma has a total of 26 billion parameters, but only 3.8 billion parameters are active during execution (due to the MoE = Mixture of Experts design).
In a quantized state, it runs on consumer-grade GPUs like the NVIDIA RTX 4090 or 5090 (18–24GB VRAM).
It can generate over 700 tokens per second on hardware close to a personal PC, without the need for dedicated data center servers.

The license is Apache 2.0 (the most permissive open-source license available for commercial use), and you can download the model weights from Hugging Face to use them freely.

Why it is still 'experimental'

To be honest, at this point, the output quality of DiffusionGemma is inferior to the conventional standard Gemma 4. Google has officially acknowledged this point.

The reason lies in design constraints. Because it is designed to prioritize speed and parallelism over precision, conventional sequential generation models are superior for applications that require complex reasoning or high quality. Google has clearly stated that they 'recommend standard Gemma 4 for applications where quality is the top priority.'

DiffusionGemma is suited for applications where speed is essentially important. It is intended for use cases such as inline code completion, real-time text editing, and interactive applications that require local operation.

What lies ahead once this design becomes widespread

DiffusionGemma is currently 'experimental.' However, when this design philosophy matures and quality reaches a level comparable to conventional models, the structure of AI utilization will change significantly.

1. Local high-speed AI becomes realistic.

An environment where AI running on your local PC responds in sub-seconds, rather than querying the cloud, becomes viable on consumer-level hardware.

2. The level of real-time editing support will improve.

Speed will no longer be a bottleneck for applications such as suggesting the next paragraph in real-time while writing text, or completing code instantly while typing.

3. Infrastructure cost calculations may change.

Instead of the pay-as-you-go model of cloud APIs, the cost structure of local execution will become a realistic option.

DiffusionGemma symbolizes the shift in the AI industry from an era of competing on 'model intelligence' to a design competition of 'how to run it and where to run it.' If the era of the typewriter is to end, it will be when the printing press is perfected.

Kei Kanzaki (Foresight Lab.)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Foresight Lab. delivers the forefront of AI and technology twice a week (Monday and Thursday). Please follow us so you don't miss the latest information.