On October 28, Shenzhen-based DiffuSpace (扩散智能) released and open-sourced Clover-35B-A3B-Preview, a diffusion large language model using a Mixture-of-Experts architecture with 35 billion total parameters and 3 billion active parameters, Securities Daily reported. In agentic-coding and key tasks, it outperformed peer Mercury 2.5 and the same-size autoregressive model Qwen3.6-35B-A3B, with generation speed 5 to 10 times faster in some scenarios.
Unlike mainstream autoregressive models such as GPT and Claude that generate tokens left-to-right, diffusion language models (dLLMs) perform global generation and parallel decoding, showing potential in logical planning and global consistency. Developers found Clover's median execution time of 1.52 seconds in 49 tasks versus 3.21 seconds for a Mercury/Jev combination — about 2.1 times faster.
Clover posted the highest scores among listed models on six agent benchmarks — SWE-bench Verified, Terminal-Bench2.1 and τ³ Overall reaching 73.6, 51.7 and 65.3 respectively. DiffuSpace, born from the HKU NLP lab and落地 in Shenzhen, has completed two financing rounds totaling several hundred million yuan.
The release advances an alternative paradigm to autoregressive LLMs and shows Chinese teams pushing both speed and agent capability in open-source models.
Source: Securities Daily / DiffuSpace
