Updated

Clover-35B-A3B-Preview (preview)

DiffuSpace released Clover-35B-A3B-Preview, a 35B-parameter mixture-of-experts diffusion language model with about 3B active parameters and a thinking mode. DiffuSpace reports results on coding and agent benchmarks; those are the company's evaluations.

By No Code MBA · Facts checked

Provider
DiffuSpace
Announcement date
Status at last check
Limited access or preview

What changed

DiffuSpace describes Clover as its first Clover model, using block diffusion with a 35B-parameter MoE backbone and about 3B active parameters per token. It supports a thinking mode and a 128K context window. Benchmark and throughput figures in the launch post are DiffuSpace-reported.

Access and availability

Download the model weights from DiffuSpace's Hugging Face model repository. The model card documents local inference and links to DiffuSpace's serving implementations.

Official access documentation

What you can build

If evaluating the preview for agent work, compare end-to-end task success and total latency on your own fixed tasks, not just tokens per second.

Limitations

The release is explicitly a preview. The model card describes a text-only model with a 128K context limit and estimates about 69 GB for BF16 weights, plus inference state. DiffuSpace's benchmark and speed results are provider-reported; long generations may become repetitive according to the model card.

Sources and release history