Diffusion Intelligence's Clover-35B takes diffusion models into agent work

Diffusion Intelligence, a team with roots in diffusion-language-model research at the University of Hong Kong, released Clover-35B-A3B-Preview on October 28. It is a diffusion language model rather than an autoregressive one, and the pitch is speed on agentic work: writing code, operating web pages, calling tools and generating HTML through an interface developers already know from chat models.

The model is a mixture-of-experts design with 35 billion total parameters and roughly 3 billion active per token, and it supports a thinking mode. It extends the team's earlier Dream 7B, a 7-billion-parameter diffusion model from April 2025, to a scale where it can be compared with production autoregressive systems.

The headline comparison is a speed test. On WindTunnel, a web-agent benchmark of 49 tasks with three attempts each, Clover's median execution time was 1.52 seconds against 3.21 seconds for a Jev plus Mercury 2.5 combination, with a slightly higher success rate — 98.6% versus 95.9%. In the demos, a solar-system simulator appeared in four to five seconds after about one second of thinking, and a LaTeX resume built from a pile of CVs took roughly ten seconds.

Agent behaviour is where the release tries to answer the obvious objection. Diffusion models can generate many positions in parallel, but they have historically been shaky at tool use: malformed calls, repeated attempts at the same step, no adjustment after a tool returns. Clover's demonstrations chain six tasks through WebMCP — invoicing, course registration, summing amounts, checking event status, and customer and invoice lookups — and complete a shopping flow that ends in a test checkout, running visibly faster than Gemini 3.6 Flash in the provided video.

The published numbers span both sides of that claim. On tool use it reports 65.3 on τ³ Overall and 67.9 on BFCL-v4. On code it reports 73.6 on SWE-bench Verified and 51.7 on Terminal-Bench 2.1. These are competitive rather than record-setting figures, which is why the speed and the continuity of multi-step execution carry more of the argument than the benchmark table.

The mechanism matters for anyone judging whether this generalises. Clover uses block diffusion: content is generated in blocks that advance in sequence, while positions inside a block update in parallel and attend to each other. The team restructured a gated DeltaNet so finished context is compressed into a continuously updated state for later steps, while the current block runs bidirectional attention. History readout is treated as a separate information unit in the attention computation, and a learnable block-size calibration counteracts the dilution of history as blocks grow.

The lineage is unusually well documented. The founders — HKU professor Kong Lingpeng and PhD students Gong Shansan and Ye Jiacheng — have worked on diffusion language models since DiffuSeq in 2022, which moved diffusion from continuous signals to sequence-to-sequence text. Dream 7B followed in April 2025; Clover is the step from generating plausible text to completing, and completing quickly, the kinds of tasks agents are judged on.

What comes next is explicitly scoped: a multimodal version that combines the diffusion language model with multimodal understanding and action generation, on-device adaptation and validation with partner Acrab, larger models aimed at long-horizon stability, and inference optimisation, quantisation and deployment work for edge hardware. The open question is whether the speed advantage survives harder, longer tasks and less curated tool environments — the preview label suggests the team knows that case is not yet closed.