According to CoinMeta, DiffuSpace transformed Alibaba’s Qwen3.6-35b-a3b into a diffusion model. Kong Lingpeng, a professor at the University of Hong Kong, founded Diffusion Intelligence and released and open-sourced clover-35b-a3b-preview. This model has a total of 35 billion parameters, with approximately 3 billion parameters activated during each inference. clover generates text using a block diffusion method and completes the output through parallel prediction and repeated refinement. According to official evaluations, clover achieved the highest scores in six agent evaluations; swe-bench reached 73.6%, and terminal-bench reached 51.7%. In terms of speed, clover generates text at a precision of bf16 at a speed of 680 token per second on clover, and 515 token per second on H200, surpassing the same-scale Qwen models (365 token per second) that use dflash2 for acceleration. The team also deployed a quantized version of int4 on RTX 5070 Ti, where the test speed reached 101.8 token per second.