Clover-35B: Diffusion language model claims twice the agent speed of Jev

What this is about

The startup Diffusion Intelligence (扩散智能) unveiled the language model Clover-35B-A3B-Preview on October 29, 2026. It belongs to the family of diffusion language models. You may know the principle from image and video generators. Here it is applied to text, code and agent tasks. According to a report by QbitAI, using it feels like using an ordinary language model, only much faster. The authors report that an interactive solar system simulator was ready after about one second of thinking and 4–5 seconds of generation. A LaTeX résumé built from raw details took about ten seconds.

Note that the claims come from the vendor or from an upbeat first-look report. According to the source, no independent tests exist yet.

The tech: how a diffusion model generates text

Ordinary ("autoregressive") language models predict the next token from the text so far. A token is a chunk of text the model works with. They write one token after another, a bit like a word-chain game. Diffusion models instead update several positions at once and refine the result over several rounds. Masked or noisy content is gradually "denoised" into fitting text.

Clover uses block diffusion. The output text is split into blocks that are produced one after another. Within a block, several positions are computed in parallel and can take each other into account (bidirectional attention, meaning attention in both directions). The catch: generating in parallel doesn't mean one round is enough. Too few rounds leave the result unfinished, and more rounds cost compute time. With code and agents, the context also keeps growing, including requirements, past actions and tool results. Storing and reading that history becomes an efficiency problem.

Clover's solution separates memory from the present. The team adapted the Gated Delta Network (GDN) for block diffusion. Context that has already been generated is compressed into a continuously updated state that later steps read from. The current block uses bidirectional attention. The readout of the history is fed into attention as its own unit of information alongside the current tokens, so the model itself learns how to weigh past against present. A learnable calibration of the block size is meant to keep the history from being "diluted" with larger blocks.

Model and performance figures

Clover is a mixture-of-experts (MoE) model. It consists of many specialized subnetworks, only some of which work on each token. In total it has 35B parameters (roughly, the dials learned during training), with about 3B active per token. Its predecessor Dream 7B had 7B parameters. According to the report, speed doesn't drop despite the size. At equal size, Clover is said to beat autoregressive models.

In the WindTunnel web-agent benchmark (49 tasks, 3 attempts each), median execution time was 1.52 seconds. The Jev + Mercury 2.5 combination needed 3.21 seconds, more than twice as long. The success rate was 98.6% for Clover versus 95.9%. Compared with Gemini, Clover is said to be "many times" faster. In one demo, it clearly beat Gemini 3.6 Flash on speed. For context, another source describes Jev as a lightweight decision model for agent workflows (model routing, safety checks, ranking), not a general-purpose language model. So the comparison is closer to one between whole systems than between models.

Agent demos

Via WebMCP (a way to make web pages operable for agents), Clover worked through six tasks in sequence: creating an invoice, signing up for a course, summing amounts, querying an event's status, and looking up customer and invoice information. In a shop scenario, it found a black T-shirt in size L, added it to the cart, entered guest details and ran through a test checkout.

That is notable because diffusion language models have so far been seen as unreliable at agent tasks. Tool-call formats went wrong, and models repeated the same step instead of reacting to feedback. Clover is said to handle the loop of calling a tool, reading the feedback and carrying on until the goal is reached.

The team and what's next

Diffusion Intelligence was founded by Prof. Kong Lingpeng (University of Hong Kong) and his PhD students Gong Shansan and Ye Jiacheng. In 2022, they introduced DiffuSeq, diffusion for text-to-text tasks. In April 2025, the core team was involved in Dream 7B, which matched equally sized autoregressive models on general knowledge, math and code. Plans include a multimodal Clover version, combining it with action generation, larger models and more stable long-running tasks. They also plan inference optimization, quantization and adaptation for edge devices. A collaboration with Acrab on hardware/software co-design is under way. To try it, the report points to a registration page (diffuspace.ai).

Our take

Diffusion language models are a counter-proposal to the standard approach. Competition comes from ELF (with He Kaiming) and Google's DiffusionGemma, among others. Clover is an early attempt to bring this approach to agents and programming, with numbers that sound strong for a "Preview". Open questions remain. There is no independent verification, and the speed comparisons use selected scenarios. The source shows neither a license nor a download for the weights, so we don't list it as an open-weights model. If you want to build agents, watch whether the latency advantage holds up in real, longer tasks.