What is available
Diffuspace’s October 30 announcement introduces Clover-35B-A3B-Preview, a text-only block-diffusion model derived from Qwen3.6. Its public checkpoint has 35 billion total parameters and about three billion active, with an Apache 2.0 license. [1][2][4]
The model card estimates about 69 GB for BF16 weights before inference state. It includes custom model code and warns that very long generations may repeat and backend numerical differences can change outputs. [2]
Check the actual setup
The project README still says official SGLang/llama.cpp code and setup guides will be added; supervised-fine-tuning code is also forthcoming. The model page lists no deployed Inference Provider, while the announcement invites applications for online access. [1][2][3]
Our take
For experienced technicians, check the full hardware budget and the supported setup before planning a local trial. Compare complete, correct tasks under matching settings; token speed alone does not establish reliability. We reviewed sources without downloading or running the model.