Alongside the base model, DiffuSpace on October 28 introduced a decision model built on Clover-35B-A3B-Preview, aimed at fast, structured judgment tasks, Zhidx reported. In a "joint mode" that reasons globally across linked questions, the model answered 95.0% correctly across 500 groups and 2,420 associated questions, fully correct in 80.4% of groups.
By contrast, the same model answering independently scored 89.8% and 62.4%, while the open-source decision model Jev historically scored 63.6% and 15.8% on the same questions. In a comparison over a contract's judgment points, Clover's joint mode got all five correct, the independent mode four, and Jev two. The decision model also defeated Jev in a game of Chinese chess, playing Black and winning.
Rather than single-point evaluation, the joint mode coordinates multiple judgments in one generation to avoid logic conflicts, enabling faster and more stable inference. DiffuSpace said a multimodal model based on Clover will follow.
The results highlight a Chinese team differentiating not just on raw model size but on inference architecture for structured, multi-step decisions.
Source: Zhidx / DiffuSpace
