Diffuspace ships 350B-parameter Clover, a diffusion model built for agent work

Diffuspace released Clover-35B-A3B-Preview on October 29. According to QbitAI, the mixture-of-experts model has 35 billion total parameters with about 3 billion activated per token and supports a thinking mode; in practice it is used much like a conventional large language model, writing code, operating web pages, calling tools and generating HTML, but it produces text through diffusion instead of next-token prediction.

Speed is the most visible difference. On the 49 tasks of the WindTunnel web agent benchmark, with three attempts per task, Clover's median execution time was 1.52 seconds against 3.21 seconds for the Jev plus Mercury 2.5 combination, with success rates of 98.6% and 95.9%. In demos it spent about a second thinking and four to five seconds producing a working solar-system simulator, and turned a batch of resumes into LaTeX in around ten seconds.

Multi-step execution is the more important claim. Clover chained six tasks through WebMCP covering invoicing, course registration, amount aggregation, event status lookup, and customer and invoice queries. In a shopping scenario it located a black T-shirt in the requested size, added it to the cart, filled guest details and completed a test checkout. Earlier diffusion language models often broke tool-call formats or looped on the same step without adapting to returned results; the demo is built around calling tools, reading feedback and continuing until the task is done.

Benchmarks cover both tool use and code: 65.3 on the overall τ³ score and 67.9 on BFCL-v4 for tool calling, 73.6 on SWE-bench Verified for real repository issues, and 51.7 on Terminal-Bench 2.1.

The architecture follows the parallel-generation premise of diffusion language models, which update many positions at once and refine over several denoising rounds rather than predicting one token at a time. Clover uses block diffusion, advancing block by block while positions inside a block update in parallel and attend to each other. To handle long agent contexts, the team rebuilt its gated Delta network so generated history is compressed into a continuously updated state, the current block uses bidirectional attention, history readout enters attention as a separate information unit, and a learnable block-size calibration limits dilution as blocks grow.

This is a scale jump rather than a new direction. The team proposed DiffuSeq in 2022 for sequence-to-sequence text generation and reached the 7-billion-parameter Dream 7B in April 2025; Clover takes the line to 35 billion total parameters. Diffuspace was founded by University of Hong Kong professor Kong Lingpeng with doctoral students Gong Shansan and Ye Jiacheng, and its work has shifted from how to generate text to how to complete complex tasks correctly and quickly.

Next on the roadmap is a multimodal version that combines the diffusion language model with multimodal understanding and action generation, plus inference optimization, quantization and deployment work for on-device use; a partnership with Acrab covers on-device system adaptation and hardware-software co-validation. Larger models and more stable long-horizon execution are also stated goals.

For the field, Clover is a test case for whether diffusion can compete with autoregressive models where latency matters most: agent loops. The real check will be independent reproduction and production deployment, not a single demo.