Beyond GPT architecture: Why Google’s Diffusion approach could reshape LLM deployment

Last month, as well as a complete suite of New AI tools and innovations, Google Deepmind revealed Gemini dissemination. This experimental research model uses an approach based on diffusion to generate text. Traditionally, large language models (LLM) like GPT and Gemini himself relied on self-regression, a step-by-step approach where each word is generated on the basis of the previous one. Diffusion language models (DLM)Also known as large diffusion language models (DLLM), take advantage of a more commonly observed method in the generation of images, starting with random noise and gradually refine them into a coherent exit. This approach considerably increases generation speed and can improve consistency and consistency.

Gemini dissemination is currently available in the form of an experimental demo; Register for the waiting list Here to have access.

(Publisher’s note: we will unpack paradigm changes such as dissemination -based linguistic models – and what it takes to execute them in production – to VB transformJune 24-25 in San FranciscoAlongside Google Deepmind, LinkedIn and other business AI managers.)

Understanding the VS self -regression

Dissemination and self -regression are fundamentally different approaches. The self -regressive approach generates sequential text, with predicted tokens one at a time. Although this method ensures strong consistency and context monitoring, it can be intensive and slow in calculation, in particular for long -form content.

The diffusion models, on the other hand, begin with a random noise, which is gradually managed in a coherent output. When applied to the language, the technique has several advantages. Text blocks can be treated in parallel, potentially producing segments or whole sentences at a much higher rate.

The broadcasting of Gemini can generate 1,000 to 2,000 tokens per second. On the other hand, Gemini 2.5 Flash has an average output speed of 272.4 tokens per second. In addition, production errors can be corrected during the refining process, improving precision and reducing the number of hallucinations. There may be compromises in terms of fine grain precision and control at the tokens; However, increased speed will change the situation for many applications.

How does the generation of text -based text work?

During the training, the DLMS works by gradually corrupting a sentence with noise on many stages, until the original sentence is made completely unrecognizable. The model is then formed to reverse this process, step by step, reconstructing the original sentence from increasingly noisy versions. Thanks to iterative refinement, he learns to model all the distribution of plausible sentences in training data.

Although the specifics of the dissemination of Gemini has not yet been disclosed, the methodology of training typical of a diffusion model implies these key stages:

Forward broadcast: With each sample of the training data set, noise is gradually added to several cycles (often 500 to 1,000) until it becomes indistinguishable from random noise.

Reverse dissemination: The model learns to reverse each stage of the NooNg process, essentially learning to “find” a corrupt sentence one step at a time, finally to restore the original structure.

This process is repeated millions of times with various samples and noise levels, allowing the model to learn a reliable delating function.

Once formed, the model is able to generate completely new sentences. DLMs generally require a condition or entry, such as a prompt, a class label or integration, to guide the generation to the desired results. The condition is injected into each step of the speeding process, which shapes an initial noise drop in structured and coherent text.

Advantages and disadvantages of models based on diffusion

In an interview with VentureBeat, Brendan O’Donoghue, scientific researcher at Google Deepmind and one of the managers of the Gemini dissemination project, has developed some of the advantages of techniques based on dissemination compared to self -regression. According to O’Donoghue, the main advantages of distribution techniques are as follows:

• Lower latencies: Diffusion models can produce a sequence of tokens in much less time than self -regressive models.
• Adaptive calculation: The diffusion models converge on a sequence of tokens at different rates depending on the difficulty of the task. This allows the model to consume fewer resources (and to have lower latencies) on easy and more tasks on more difficult resources.
• Non -causal reasoning: Due to bidirectional attention in Denoisi, tokens can take care of future tokens in the same generation block. This allows non -causal reasoning to occur and allows the model to make global modifications in a block to produce more coherent text.
• Iterative refinement / self-correction: The storage process involves sampling, which can introduce errors as in self -regressive models. However, unlike the self -regressive models, the tokens are brought back to the Dencher, which then has the possibility of correcting the error.

O’Donoghue has also noted the main drawbacks: “higher cost of service and a slightly higher toker time (TTFT), as self -regressive models will immediately produce the first token. For broadcasting, the first token can only appear when the whole sequence of token is ready. ”

Performance benchmarks

Google says that gemini broadcast performance is Comparable to Gemini 2.0 Flash-Lite.

Reference	Type	Gemini dissemination	Gemini 2.0 Flash-Lite
Livecodebench (V6)	Code	30.9%	28.5%
Bigcodebench	Code	45.4%	45.8%
LBPP (V2)	Code	56.8%	56.0%
Swe-bench checked *	Code	22.9%	28.5%
Human	Code	89.6%	90.2%
Mbpp	Code	76.0%	75.8%
GPQA diamond	Science	40.4%	56.5%
Likes 2025	Mathematics	23.3%	20.0%
Extra hard Big-Banc	Reasoning	15.0%	21.0%
MMLU Global (Lite)	Multilingual	69.1%	79.0%

* Non -aged evaluation (single turn modification only), maximum prompt length of 32K.

The two models were compared using several landmarks, with scores based on the number of times the model produced the correct answer during the first try. Gemini dissemination worked well in coding and mathematics tests, while Gemini 2.0 Flash-Lite had the advantage of reasoning, scientific knowledge and multilingual capacities.

As the dissemination of Gemini evolves, there is no reason to think that its performance will not catch up with more established models. According to O’Donoghue, the gap between the two techniques is “essentially closed in terms of reference performance, at least with relatively small sizes that we have increased. In fact, there may be a performance advantage for dissemination in certain areas where non -local consistency is important, for example, coding and reasoning. ”

Testing gemini dissemination

Venturebeat had access to the experimental demo. By putting the broadcasting of Gemini to the test, the first thing we noticed was speed. During the execution of the suggested prompts provided by Google, including the construction of interactive HTML applications such as the Xylophone and the planet Tac Toe, each request finished in less than three seconds, with speeds ranging from 600 to 1,300 tokens per second.

To test its performance with a real world application, we asked Gemini Diffusion to create a video chat interface with the following prompt:

Build an interface for a video chat application. It should have a preview window that accesses the camera on my device and displays its output. The interface should also have a sound level meter that measures the output from the device's microphone in real time.

In less than two seconds, Gemini Diffusion created a work interface with a video overview and an audio counter.

Although this is not a complex implementation, it could be the beginning of a MVP which can be completed with a little additional incentive. Note that Gemini 2.5 Flash has also produced a work interface, although a slightly slower rhythm (about seven seconds).

GEMINI DIFFUSION also has “instant modification”, a mode where text or code can be stuck and modified in real time with minimum incentive. Instant edition is effective for many types of text editing, including grammar correction, updating the text to target different readers’ characters or the addition of key referencing words. It is also useful for tasks such as refactoring code, adding new features to applications or conversion of an existing code base into a different language.

Business use case for DLMS

It is sure to say that any application that requires rapid response time is to benefit from DLM technology. This includes real-time and low latency applications, such as conversational AI and chatbots, live transcription and translation, or assistants of semi-automatic entry and COD.

According to O’Donoghue, with applications that take advantage of “online edition, for example, taking a text and making changes in place, the diffusion models are applicable in a way that the self -regressive models are not.” DLMs also have an advantage with problems of reason, mathematics and coding, due to the “non -causal reasoning granted by bidirectional attention”.

DLMs are still in their infancy; However, technology can potentially transform the way in which language models are built. Not only do they generate text at a much higher rate than self -regressive models, but their ability to go back and correct errors means that, ultimately, they can also produce results with greater precision.

The dissemination of Gemini enters a growing DLMS ecosystem, two notable examples being Mercurydeveloped by creation laboratories, and LladaAn open source model from GSAI. Together, these models reflect the wider momentum behind the generation of languages ​​based on diffusion and offer an evolutionary alternative and parallelizable to traditional self -regressive architectures.