Diffusion Intelligence releases Clover-35B-A3B-Preview, a 35B-parameter diffusion language model for agents

Diffusion Intelligence (扩散智能) has released Clover-35B-A3B-Preview, a diffusion language model built for coding, tool use and multi-step agent tasks. It uses a mixture-of-experts (MoE) architecture with 35B total parameters and about 3B activated per token, and it supports a thinking mode. The announcement was reported in Chinese by QbitAI (量子位) on 2026-10-29.

Key specifications

• Parameters: 35B total, about 3B active per token (MoE)
• Generation method: block diffusion, with thinking mode
• Predecessor: Dream 7B (7B parameters), released in April 2025 with participation from the team's core members
• Context window: not stated in the report
• Pricing: not yet disclosed
• Access: registration at diffuspace.ai/register

Benchmarks (published by the team)

All figures below come from the team's own published data.

Benchmark	Clover score
SWE-bench Verified	73.6
Terminal-Bench 2.1	51.7
τ³ Overall	65.3
BFCL-v4	67.9

The team says the τ³ Overall and BFCL-v4 results place Clover among the top entries in tool-use evaluations.

On WindTunnel, a web-agent benchmark of 49 tasks with 3 attempts per task, Clover's median execution time was 1.52 seconds. The Jev + Mercury 2.5 combination recorded 3.21 seconds. Attempt success rates were 98.6% for Clover and 95.9% for Jev + Mercury 2.5. According to the report, Clover was several times faster than Gemini in the same table. The team also claims Clover outpaces autoregressive models of the same size.

Architecture

Autoregressive models predict one token at a time. Diffusion language models update multiple positions in parallel and refine them over several rounds. Clover uses block diffusion: content is split into blocks that are generated in sequence, while tokens inside a block update in parallel and attend to each other.

The team rebuilt the Gated Delta Network (GDN) around block diffusion. Previously generated context is compressed into a continuously updated state that later steps read from. The block being generated uses bidirectional attention. The memory readout is treated as a separate information unit that takes part in attention alongside the current tokens, so the model learns how to weigh history against current content. A learnable block-size calibration is meant to reduce dilution of historical information as blocks grow.

Agent demonstrations

In official demos, Clover used WebMCP to complete six consecutive tasks: invoicing, course registration, amount summation, event status lookup, customer information lookup and invoice information lookup. QbitAI reports it ran faster than Gemini 3.6 Flash in the comparison video. In an online shopping scenario, it located a black, size L T-shirt, added it to the cart, filled in guest details and completed a test checkout.

QbitAI's hands-on tests included a solar system simulator, which the model produced after about 1 second of thinking and 4 to 5 seconds of generation. A LaTeX résumé took roughly ten seconds.

QbitAI notes that earlier diffusion language models were unstable on agent work, with tool-call format errors and repeated attempts at the same step without adjusting to returned results.

Roadmap and team

Diffusion Intelligence was co-founded by Professor Kong Lingpeng of the University of Hong Kong and his PhD students Gong Shansan and Ye Jiacheng. The group's diffusion language work dates to DiffuSeq in 2022.

The team says a multimodal version of Clover will follow. It is working on combining diffusion language models with multimodal understanding and action generation, and it is validating on-device deployment. Its collaboration with Acrab covers system adaptation and hardware-software co-validation for on-device use. Planned work includes larger models, better long-horizon stability, inference optimization, quantization and deployment adaptation.

What this means

Clover moves the team's diffusion language model from 7B parameters in Dream 7B to 35B total parameters, and it targets agent workloads: tool calls, reading feedback and continuing until a task is done. The reported WindTunnel numbers pair a lower median execution time with a higher success rate than the Jev + Mercury 2.5 combination. The SWE-bench Verified, Terminal-Bench 2.1, τ³ and BFCL-v4 scores are team-published and come without comparison tables in the report. The model carries a Preview label, and the context window and pricing are not yet disclosed.