gensai.aiLLMs (Large Language Models) such as GPT and Claude revolutionised text generation. But a new generation of models is starting to take shape: DLLMs, or Diffusion Large Language Models. Drawing on image-generation techniques, these models reduce hallucinations, speed up processing and handle context better. An advance that could transform how generative AI is used in professional environments.
A technical break: diffusion applied to language
Unlike traditional LLMs, which generate text word by word, DLLMs take an entirely different approach, borrowed from image-generation models. The principle draws on diffusion models, like those that made AI image creation possible (for example Stable Diffusion).
How does it actually work?
- The model starts from a random sequence of tokens (pieces of text), the equivalent of a "noisy image".
- It then progressively denoises this sequence to produce coherent text, across several successive steps, each guided by the context and the objective (the prompt).
This generation process is not linear but global, which makes it possible to optimise overall coherence. Each iteration takes the whole into account, rather than building sentence by sentence.
Three major advantages for business uses
1. Fewer hallucinations
The main benefit of DLLMs is reducing factual errors, which are frequent with classic LLMs. By generating text as a whole through an iterative refinement process, DLLMs ensure better logical coherence and reduce the risk of producing invented information.
2. Faster generation
Unlike autoregressive models (LLMs), which must wait for the previous word before generating the next, DLLMs can parallelise their generation. This considerably reduces latency and speeds up processing, particularly for long or complex answers.
3. Better control over the output
The different generation steps make it possible to intervene more precisely in the process, by modifying or readjusting certain parameters. This modularity eases business integration, particularly in sensitive environments (legal, healthcare, customer relations and so on).
Towards lighter, better-suited models
DLLMs are part of a broader trend towards lighter architectures, optimised inference costs and easier local or sovereign-server deployments. This movement accompanies the search for more frugal, more reliable and more customisable solutions, in line with regulatory (GDPR, AI Act) and industrial requirements.
A young technology, but a promising one
The first results presented in publications from Google DeepMind and Anthropic show that this diffusion approach applied to text is not only effective but potentially disruptive for professional use cases — particularly those requiring speed, precision and contextualisation at once.
In summary
| Characteristic | Classic LLMs | DLLMs (Diffusion LLMs) |
|---|---|---|
| Type of generation | Word by word (autoregressive) | Complete sequence (diffusion) |
| Speed | Slower (sequential) | Faster (parallelisable) |
| Accuracy | Prone to hallucinations | Fewer errors, more coherence |
| Control | Limited | Fine-grained tuning possible |
DLLMs do not replace LLMs yet, but they are already sketching the outlines of a new generation of generative AI — better suited to real-world requirements: efficiency, reliability and frugality.