The mainstream model is still expanding along the path of return, but another technological path is accelerating near the industrial milestone.
A joint test provides new evidence for this possibility.
On September 1, a team of distributed language models, DiffuSpace, and Acrab, an Asian intelligent computing platform, demonstrated that dLLM can increase the speed of the Agent by about five times in an edge-side scenario. Based on this test, the parties plan to accelerate the deployment of distributed language models in scenarios such as AI PCs, smart cars, robots, and smart homes.
The Diffusion Language Model (dLLM), with its uniqueness of "universal generation", is seen as one of the key directions for the next generation of general AI base models, and is even predicted to replace or replace the GPT class self-return model within two years.
Different from the mainstream model of the GPT class, the dLLM model takes advantage of the diffusion mechanism of the image model to create a holistic content framework, then combine the results of continuous modification in the context, both to form more holistic consistency, and to achieve 5 times or even 10 times the acceleration of reasoning.
This difference will be further amplified on the edge of AI.
The generating paradigm of the traditional self-recurring model is that each time a chip generates only one token chip, it is not possible to efficiently use the parallel computing power of the end-side chip; however, a distributed language model can assign multiple locations simultaneously to GPU processing, converting the parallel computing capability of the chip into real speed.
The Agent scenario further amplifies the importance of generating speed. A complete Agent task, often including planning, retrieving, calling tools, executing and calibrating, may be repeatedly called upon by the model. Delays caused by each generation add up to the entire task chain.
In the case of an AI PC, a local agent may need to read emails, schedule, retrieve files and generate follow-up plans at the same time, and a distributed model can run parallel to multiple related steps and synchronize multiple interrelated tasks according to changes in real time.
Distributed intelligence DiffuSpace is one of the world's earliest teams to engage in dLLM paradigm research, co-founded by Prof. Kunming Liu, co-director of the National Laboratory of Linguistics in Hong Kong, and his Ph.D. students, Dr. Suzan Tsai and Dr. Yatsui. The core team has long been deeply engaged in the study of distributed language models, and is at the international forefront in this field, with world-leading breakthroughs in the three main dLLM core problems of conditional generation, discrete signal modeling, and scaling.
In 2022, the team launched DiffuSeq, pioneering the technological path of distributed model processing sequence to sequence text generation; in 2025, the team launched the Dream 7B model for the first time fully exceeding the same parameter-scale self-return model and outperforming the 671B's DeepSeek V3 in planning capabilities.
Around this technological route and industrial applications, we had an exchange with the distributed intelligent DiffuSpace.
At the heart of the machine: Why could the edge be the scene where a distributed language model is the first to create an advantage?
DiffuSpace: The acceleration advantages of a distributed model are that it is relatively regressive, and is most evident when the model only handles one user's task. End-side reasoning, such as AI in mobile phones, computers and cars, usually serves only one local user; at the same time, end-side users are very sensitive to response speeds and expect the model to complete tasks in a timely manner.
The heart of the machine: where exactly did the speed advantages of the spreading language model come from?
DiffuSpace: Self-return models typically generate only one Token forward at each step. A spreadsheet model can generate multiple Tokens at the same time, closer to the process people think about. When people think about problems, they don't necessarily write all the tokens, all the words one word at a time in their minds.
Now, when using a GPT-class model, users may have to wait a minute; if the speed increases to ten times, they only have to wait ten seconds. Once they experience this sense of rapid response, it is difficult to return to the previous fast-thinking experience. In some scenarios, speed is very important. For example, two functions can be relatively independent modules, brackets, definitions, and other structures with higher predictability. Spread models can usually generate 7 to 10 tokens at each step. This will directly increase the generation speed and further improve service efficiency and lower unit costs.
Machine Heart: In addition to generating faster, how will dLLM change the user experience of Agent?
DiffuSpace: In the example of the Coding Agent, if the user modifies the content of the previous rounds of conversations, the auto-return model can only be generated backwards from the modified location, and the resulting content is difficult to continue using. The spread model can adapt to the modification and adjust the subsequent content accordingly, as much as possible to retain the still valid part, only changing the content of the photo being affected. The user will intuitively feel that the model is more editable, flexible, and intelligent.
In the longer term, a distributed language model can achieve higher levels of intelligence, breaking certain human cognitive boundaries, and moving towards AGI, ASI. Distributed language models are better at global searching and better at diversity exploration. Both are very beneficial for the RSI.
In the face of unknown problems, the model needs to pave as many different paths of exploration as possible, and then filter the answers through reliable verifiers to be better able to simultaneously generate diverse candidate schemes, thus potentially increasing the likelihood of finding the right solution. At the same time, it is able to combine overall goals and follow-up steps to adjust current decisions rather than just going forward based on the content that has already been generated, thus reducing the situation of being trapped in local optimization.
The heart of the machine: Does higher reasoning efficiency mean higher costs for the training phase?
DiffuSpace: The spread model can be understood as "training for the worst, testing for the best". When training, it needs to see more noise and position combinations, and may need to train more rounds on the same data; but at the reasoning stage, higher parallels can be achieved. The model only needs to be trained once, but serves a large number of users, so it is generally cost-effective to use more training in exchange for long-term reasoning efficiency.
At this stage, we can also make direct use of the parameters of existing regression models through adaptation, reducing the cost of scaling the scattering model. The language knowledge and capabilities that the regression model has learned can be inherited by the more general scattering model and then added to the bi-directional, global information.
At the heart of the machine: What are the other areas where dLLM is moving away from large-scale industrial applications?
DiffuSpace: It's not an easy thing to really make something big, to do it right. From models, data, computing, to engineering and ecology, everything is challenging, and we're moving forward one step at a time. For example, the return model already has a lot of mature algorithms and infrastructure like FlashAttention, while the spreadsheet language model's algorithms, reasoning systems and infra are not so perfect.
However, a more positive change is that more people are entering this field, we are not fighting alone, there have been many open source projects and everyone is contributing to this field. I believe that the more people do this, the better the whole community and the ecology will improve step by step.
At the heart of the machine: Why is it possible that a distributed language model could replace the GPT class self-return model within the next two years?
DiffuSpace: The widespread language model is not yet complete, partly because it has not been funded and researched long enough. Any technological path from academic exploration to large-scale applications requires a process of accumulation.
But the distributed language model has gained increasing attention and has gradually taken on a place of its own. In the next six months, we will continue to expand the model capacity and influence, transforming more research advantages into real models and products; in six months to a year, we will also see more vertical products. In the fourth quarter of this year, we will release a new distributed language model, with plans to expand the size of the model parameters to a quantitative level next year.
Within the next one or two years, the proliferation language model is likely to replace the current self-return model. Of course, this is a very radical goal, and there is much more to be done.