DiffuSpace adapted Qwen3.6 into a diffusion model, delivering 680 tokens per second on a single GB300 card.

Beating AI News Flash: DiffuSpace, founded by University of Hong Kong professor Lingpeng Kong, has released and open-sourced the Clover-35B-A3B-Preview model. Modified from Alibaba's Qwen3.6-35B-A3B, the model has 35 billion total parameters, with around 3 billion activated per inference. Traditional large models generate tokens sequentially left-to-right, while Clover uses block diffusion: it generates small text segments at a time, then completes output via parallel prediction and iterative refinement. The team also modified Gated DeltaNet to compress and store historical information, while enabling bidirectional interaction between tokens in the current text block, balancing generation speed and contextual understanding. In six official agent benchmarks, Clover scored the highest among all listed models. Its SWE-bench Verified score reached 73.6%, Terminal-Bench 2.1 hit 51.7%, outperforming the original Qwen3.6-35B-A3B, as well as diffusion models like 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 runs at 680 tokens per second on a single NVIDIA GB300 with BF16 precision and a single concurrent internal agent task. On H200, it hits 515 tokens per second, outperforming the same-scale Qwen model accelerated with DFlash2 (365 tokens per second). The team also deployed the INT4 quantized version to an RTX 5070 Ti with only 16GB of VRAM, achieving 101.8 tokens per second on the MATH-500 benchmark.