正文 · AI 翻译
How can a diffusion model work as an agent?
By henry, reporting from Aofeisi
QbitAI | WeChat public account QbitAI
By henry, reporting from Aofeisi
QbitAI | WeChat public account QbitAI
I give up! Even surfing the front lines of AI every day, I can still fall behind the latest developments.
Who would have thought that “diffusion models,” which I associated with generating images and videos, could also write code, generate web pages, call tools, and act as agents!
Just now, Diffusion Intelligence released a brand-new diffusion language model—
Clover-35B-A3B-Preview.
Although its name adds the word “diffusion,” in actual use, whether writing code, operating web pages, calling tools, or generating HTML, it works much like the large language models we know.
My first impression from using it can be summed up in one word: “fast.”
For example, when I asked it to generate a solar system simulator, Clover thought for about 1 second, and then, in just 4–5 seconds, an interactive solar system simulator appeared.
As another example, I gave it a pile of résumé information and asked it to turn it into LaTeX résumé code. It finished in about 10 seconds.
Honestly, anyone who has used GPT probably knows what it feels like to stare at “thinking” for ages. Otherwise, there would not have been so much interest lately in Jev, which emphasizes quick decisions.
Speaking of Jev, Clover actually went head-to-head with it on speed this time.
Across 49 tasks in the WindTunnel web agent benchmark, with 3 attempts per task, Clover's median execution time was 1.52 seconds, versus 3.21 seconds for the Jev + Mercury 2.5 combination—less than half as long.
Its attempt success rate was also slightly higher: 98.6% versus 95.9%, respectively.
Compared with Gemini in the same table, it was several times faster!
Well, well, well, diffusion model—are you trying to drive autoregressive models crazy? (doge)
Back to the model: Clover-35B-A3B-Preview uses an MoE architecture, with 35 billion total parameters and about 3 billion parameters activated per token, and supports a thinking mode.
It takes the team's diffusion language models from the 7 billion parameters of the earlier Dream 7B to a total of 35 billion parameters.
More crucially, Clover did not slow down as its parameter count grew; in comparisons with models of the same size, it outperformed autoregressive models on speed.
Beyond the interactive page and code generation shown earlier, Clover also demonstrated agent capabilities for complex tasks and WebUI generation in the official demos.
Using WebMCP, Clover can handle 6 tasks in succession involving invoicing, course registration, summing amounts, checking event status, and looking up customer and invoice information.
As the comparison video below shows, Clover completed the tasks successfully and left Gemini 3.6 Flash far behind in execution speed.
In an online shopping scenario, Clover can also find a T-shirt in the specified style, in black and size L, add it to the cart, fill in guest information, and complete the test checkout process.
Honestly, its handling of these consecutive tasks is what makes this demo worth watching.
Diffusion language models may have the potential for speed through parallel generation, but their performance on agent tasks has previously been insufficiently stable:
The model may get the tool-call format wrong or repeatedly try the same step without adjusting its actions based on the returned results.
Clover's progress this time shows in these successive operations: calling a tool, reading the feedback, and continuing until the task is done.
As for benchmarks, according to figures released by the team, Clover scored 65.3 on τ³ Overall and 67.9 on BFCL-v4, both tool-use evaluations, placing it among the leaders.
Having looked at agent tasks, let's look at code.
I also asked Clover to generate a web interface for a personal agent. It looked quite convincing, and the whole generation process took only a little over 10 seconds.
Clover also posted a good set of results on code benchmarks.
It scored 73.6 on SWE-bench Verified, which tests the ability to solve problems in real code repositories, and 51.7 on Terminal-Bench 2.1, which tests the ability to carry out terminal tasks.
From generating interactive pages and calling tools to completing business workflows and taking code benchmarks, Clover, as a diffusion language model, has indeed taken on quite a few agent tasks this time.
Now this is interesting. How does the diffusion approach I associate with denoising images generate text and write code when applied to language models?
Could the demos above really have been produced by AIGC?
Generally speaking, the autoregressive (AR) language models we know predict the next token from the preceding text, generate it, then generate the next one—the process often described as a word-chain game.
Here, a token can be understood simply as a piece of text that the model processes.
Unlike the one-by-one, sequential prediction and generation of autoregressive models, diffusion language models generate in another way: they allow multiple positions to be updated at once, then refine them over several rounds to arrive gradually at a result.
Here, diffusion denoising can be understood as gradually restoring masked or disturbed content into text that fits the context.
Specifically, Clover uses block diffusion: it divides the content to be generated into blocks, advances from block to block in sequence, and updates multiple positions within each block in parallel so they can refer to one another.
This gives a single round of computation the chance to advance generation at multiple positions, reducing the wait caused by generating them one by one.
But being able to generate them together does not mean they can be written well in a single round.
Diffusion models often need multiple rounds of denoising to settle on a result gradually. With too few rounds, the content may not be fully refined; adding rounds also increases the computational cost.
For code and agent tasks, the context also keeps growing. Earlier requirements, actions already taken, and results returned by tools can all affect the next step, making the storage and retrieval of past information an efficiency issue as well.
Clover's approach to improvement is to give historical memory and current generation distinct roles.
It reworks the Gated Delta Network (GDN) around block diffusion: generated context is compressed into a continuously updated state for later steps to read, while the block currently being generated uses bidirectional attention so tokens within it can refer to one another.
At the same time, Clover treats the readout of historical memory as a separate unit of information that participates in attention alongside the current token, allowing the model to learn how to weigh earlier context against current content. It also uses learnable block-size calibration to reduce the dilution of historical information as blocks grow larger.
In other words, these adjustments aim to make parallel generation, the use of historical information, and computational efficiency work better together. That is also key to Clover's further move into code and agent tasks as a diffusion language model.
Beyond Clover, diffusion language models have drawn attention over the past 2 years chiefly because of 2 advantages they offer over autoregressive models:
For one, diffusion language models allow multiple positions to be generated at once, potentially reducing the sequential wait involved in token-by-token generation. This potential is especially relevant to agents that repeatedly generate code and call tools.
For another, earlier and later content can refer to each other during generation. With bidirectional modeling, multiple positions can be adjusted together through iteration, offering another way to handle dependencies between pieces of content. The interaction within Clover's current block is one implementation of this.
Researchers and companies have already tried different approaches to these possibilities.
ELF, which Kaiming He worked on this year, attempts to generate language in a continuous embedding space; Google's DiffusionGemma explores parallel text generation; and major Chinese tech companies are also researching diffusion language models for coding tasks.
The specific methods differ, but everyone is trying to turn this alternative way of generating text into practical capability and efficiency.
With Clover, the team has put its progress into code, tool calls, and consecutive tasks for users to try for themselves.
Finally, let's return to the team behind Clover—
Diffusion Intelligence.
Reportedly, Diffusion Intelligence was co-founded by University of Hong Kong professor Lingpeng Kong and his doctoral students Shansan Gong and Jiacheng Ye. The team members' exploration of diffusion language models dates back to 2022.
That year, they introduced DiffuSeq, studying how to use diffusion models for sequence-to-sequence text generation: given one passage of text, generate another based on specified conditions.
One hurdle at the time was applying diffusion methods, which excel at handling continuous signals such as images, to discrete natural language.
By April 2025, core team members had helped launch Dream 7B, advancing the research to a language model with 7 billion parameters.
According to the results released at the time, Dream 7B achieved performance comparable to autoregressive models of the same scale on general, math, and coding tasks, while also showing planning ability and flexibility in generation order.
Now, with Clover, the model has grown further, and writing code, calling tools, and executing tasks in succession have become key parts of the demonstration.
This path has led step by step from initial research into how to generate text to how to complete complex tasks correctly and quickly. The team is now also expanding into multimodal and on-device applications.
According to the introduction, a multimodal version of Clover will also be released later. The team is working to combine diffusion language models with multimodal understanding and action generation, while continuing adaptation and validation for on-device scenarios. Its collaboration with Acrab is advancing system adaptation and coordinated software and hardware validation for on-device scenarios.
According to the team's plans, it will continue exploring larger models, improve stability on long-horizon tasks, and advance on-device applications through inference optimization, quantization, and deployment adaptation.
And finally, let's use the newly released Clover to entertain a thought:
When humans think, do we work through one sentence after another, or do we first form a vague overall picture and then gradually fill in the details—
Is the brain more like an autoregressive model or diffusion?
Try it here
[1]https://diffuspace.ai/register