- Clover uses 35 billion total parameters and activates about 3 billion per token, the report says.
- QbitAI describes its own coding tests; published benchmark figures remain attributed to the team.
Diffusion moves into agent workflows
DiffuSpace has introduced Clover-35B-A3B-Preview, a diffusion language model aimed at coding, tool use and agent workflows, QbitAI reported on October 29. The team was co-founded by University of Hong Kong professor Kong Lingpeng and two of his doctoral students, according to the report.
Clover uses a mixture-of-experts architecture with 35 billion total parameters, activating about 3 billion per token, and supports a thinking mode. Its block-diffusion approach updates multiple positions within a block while proceeding through blocks in sequence.
Tests and plans
QbitAI's writer describes trying the model on an interactive solar-system simulation, LaTeX resume code and a personal-agent interface. Those examples are a limited hands-on account rather than a comprehensive reliability assessment.
The team also presents demonstrations of consecutive tool calls and test shopping workflows, alongside reported coding and tool-use benchmark results. The report says future work includes a multimodal version and deployment optimizations; these remain plans. Its hands-on examples and team demonstrations do not establish general performance across production tasks.
Sources & context
Go to the original material. Company claims remain attributed to their sources.
01
