DiffuSpace converts Qwen3.6 into a diffusion model, achieving 680 tokens/s on a single GB300 GPU.

Beating AI News Flash: DiffuSpace, founded by University of Hong Kong professor Kong Lingpeng, has released and open-sourced Clover-35B-A3B-Preview.

The model is built on Alibaba's Qwen3.6-35B-A3B, with 35 billion total parameters and approximately 3 billion activated parameters per inference.

Traditional large models typically generate tokens one by one from left to right, while Clover uses block diffusion, generating a short segment of text at a time and completing output through parallel prediction and iterative correction. The team also modified Gated DeltaNet to compress and store historical information while enabling bidirectional interaction among tokens within the current text block, balancing generation speed and context understanding.

In six officially published Agent benchmarks, Clover achieved the highest scores among the listed models. It reached 73.6% on SWE-bench Verified and 51.7% on Terminal-Bench 2.1, surpassing the original Qwen3.6-35B-A3B as well as diffusion models including LLaDA2.2 Flash and Mercury 2.5. However, the comparison list did not include Mercury 3, released on October 19.

In terms of speed, Clover achieved 680 Token/s on a single NVIDIA GB300 running internal Agent tasks at BF16 precision with single concurrency. On H200, it reached 515 Token/s, surpassing a same-scale Qwen model accelerated with DFlash2 (365 Token/s). The team also deployed an INT4 quantized version on an RTX 5070 Ti with only 16GB of VRAM, achieving 101.8 Token/s on the MATH-500 test.