DiffuSpace’s October 30, 2026 launch article introduced Clover-35B-A3B-Preview, an open-weight diffusion language model aimed at coding and tool use. Preview weights are downloadable, while unrestricted hosted API access is not established by the reviewed primary materials. Its parallel generation approach is worth evaluating against the demands of a complete agent workflow.
The short answer: downloadable weights are available, but a generally available hosted API is not established by the reviewed primary materials. The model remains explicitly named Preview, and DiffuSpace's performance figures are its own reported results, not Nerova measurements.
Downloadable weights are not the same as hosted API access
The Hugging Face repository contains actual model files, including 26 main Safetensors shards, configuration files and custom Python implementation files. Its file listing totals approximately 69.3 GB. This is more than an announcement or a promise to release weights later.
DiffuSpace's launch materials direct readers to request online access. They do not establish an unrestricted public API with documented service commitments or pricing. Separately, the Hugging Face model card says no Inference Provider is serving the model there. That statement concerns Hugging Face providers; it does not prove that no private or separately arranged DiffuSpace service exists.
For procurement, keep three statuses separate: an announcement has been published; preview weights can be downloaded; hosted API availability still needs direct confirmation. Do not budget a production integration around an assumed endpoint or an unverified token price.
What Clover changes about text generation
According to the model card, Clover is a text-only derivative of Qwen3.6-35B-A3B, with approximately 35 billion total parameters, 3 billion active parameters per token and a 131,072-token combined prompt-and-response limit. Its block-diffusion implementation starts with masked positions, fills sufficiently confident positions in parallel, then adds the completed block to the preceding context. Defaults use 32-token blocks and a 0.95 confidence threshold.
The operational implication is a possible reduction in generation time within repeated agent steps. That is a reason to evaluate Clover—not a guarantee that browser navigation, database queries, approvals or an entire business process become proportionally faster.
The speed claim is specific; deployment is backend-dependent
DiffuSpace reports 680 tokens per second on internal agent workloads using one NVIDIA GB300 GPU, BF16 precision and batch size one. Its own website also reports 515 tokens per second on an H200. These are vendor measurements under specified conditions, not universal throughput guarantees. The same launch coverage reports 73.6 on SWE-bench Verified and 51.7 on Terminal-Bench 2.1; those scores likewise remain company-reported evidence.
The official GitHub README distinguishes Transformers block-diffusion inference, SGLang block diffusion and self-speculative decoding, and llama.cpp self-speculative decoding. Those paths should not be treated as interchangeable. It recommends the official SGLang implementation for serving and evaluation, warns that numerical differences can change generated text, and says backend code and setup guides will be added to the repository. Supervised fine-tuning instructions are also marked coming soon.
The README's BF16 example requires roughly 69 GB for weights plus inference state. About 3 billion active parameters therefore does not mean a 3-billion-parameter memory footprint. Its example specifies Transformers 5.19.0 and PyTorch 2.10 or later; teams should reproduce the documented environment rather than assume a generic serving command proves compatibility.
What an agent pilot should establish next
Our assessment: start with a reversible workflow, such as preparing a support-ticket response for review, rather than granting immediate write access to customer systems. Measure successful task completion, invalid tool arguments, retries, total elapsed time and infrastructure consumption. Separate model generation from tool waiting so a token-speed improvement cannot conceal an unchanged business outcome.
Two preview limitations deserve explicit tests. The model card acknowledges repetition during very long generations. And the October 30 article describes a separate decision model as a later release, while its agent harness and desktop coding client remain under development. Those future components should not be counted as capabilities already delivered with the downloadable checkpoint.
The next useful milestones are reproducible backend setup, independently checkable workload results and clearly documented hosted access. Until then, Clover is a concrete evaluation opportunity, not evidence by itself that a business should replace its existing agent stack.