Clover’s benchmark claims: what can readers verify?
DiffuSpace has published Clover’s weights and model documentation. Its reported benchmark wins and generation speeds still need independent reproduction under disclosed test conditions.
DiffuSpace released Clover-35B-A3B-Preview on October 30, 2026, with a launch post ↗, a Hugging Face model card ↗, and a public code repository ↗. That is a useful starting point for scrutiny. It does not, by itself, establish that the model beats its competitors or reaches the reported speeds in someone else's hands.
What readers can inspect. The Hugging Face file listing exposes 26 model weight shards, configuration and model code. The model card describes a 35-billion-parameter mixture-of-experts model with about 3 billion active parameters per token, derived from Qwen3.6-35B-A3B and adapted for block diffusion. It supplies an example for loading the model. The public GitHub README describes the architecture and inference paths, but says its SGLang and llama.cpp backend code and setup guides will be added to that repository. Availability of weights is concrete evidence of a release; it is not a benchmark rerun.
What DiffuSpace reports. In its own launch post, the team gives Clover a score of 73.6 on SWE-bench Verified and 51.7 on Terminal-Bench 2.1. It also reports 680 generated tokens per second on internal agent workloads using one NVIDIA GB300, BF16 precision and batch size one. For an H200 comparison, it reports 515 tokens per second for Clover versus 365 for a DFlash2-accelerated Qwen setup. Those numbers are attributed to DiffuSpace. We have not treated the company's charts or demos as independent confirmation.
What a reader would need to test. A fair reproduction needs the exact model revision, benchmark dataset and harness, prompts and scoring rules, inference backend, decoding settings, precision, batch size, hardware and comparator configuration. Speed on one internal workload may not predict latency or quality in another. The same applies to scores from selected benchmarks: they describe particular tests, not every coding or agent task.
The immediate question is therefore narrower than “Is Clover the fastest?” Readers can inspect the released files and methods, and independent teams can try the model. The sources reviewed for this draft do not identify an independent evaluator that has reproduced the headline speed and benchmark comparisons. Until that happens, the figures remain DiffuSpace's reported results. The release is newsworthy because it gives others material to test—and the next credible update will be a transparent, repeatable test of those claims.
BEHIND THE STORY
How to read this piece
These notes were supplied by the author and reviewed for publication. They do not independently verify the claims.
Significant points
The Hugging Face listing exposes model weight files, configuration and model code; this supports the claim that a model artifact was released.
DiffuSpace reports 73.6 on SWE-bench Verified and 51.7 on Terminal-Bench 2.1; these are its own benchmark results.
DiffuSpace reports 680 tokens/s on one GB300 using BF16, batch size one and internal agent workloads, plus 515 versus 365 tokens/s in an H200 comparison.
The reviewed sources do not identify an independent reproduction of those headline scores or speeds.
Who
DiffuSpace and its Core Team released and described Clover. The model card says Clover derives from Qwen3.6-35B-A3B. DiffuSpace is the source of the benchmark and speed claims; no independent evaluator is identified in the sources reviewed.
What
DiffuSpace published Clover-35B-A3B-Preview, a block-diffusion mixture-of-experts language model, with model files and documentation. Its launch post also claims leading scores and faster generation on selected tests.
When
DiffuSpace’s Hugging Face launch post is dated October 30, 2026. This draft reviews the public release as of November 1, 2026; the source pages may change after publication.
Where
The release is on Hugging Face, with a related public GitHub repository. The reported speed tests use NVIDIA GB300 and H200 hardware; no geographic location is necessary to understand the release.
Why it matters
If reproduced, stronger agent-task performance at lower generation time could change how developers choose an open model. The publicly listed files let others test the claim, while the creator’s own measurements alone do not establish general superiority.
Why now
The model and launch post appeared on October 30, 2026. Fresh weights and documentation make this an appropriate time to distinguish an inspectable release from reported results before those figures circulate as established fact.
Claim attribution
DiffuSpace Core Team made the performance and architecture claims in its launch post. Its model card and GitHub README describe the released files and implementation. AINewsIQ’s analysis distinguishes these first-party claims from what a reader can inspect directly; it does not claim an independent benchmark run.
What remains unknown
The reviewed sources do not establish an independent rerun of the benchmark or speed comparisons. Exact revisions, harnesses, prompts, scoring rules, backend and decoding settings, comparison configuration, and workload quality need scrutiny for a fair reproduction. AINewsIQ has not downloaded or executed the model for this draft.
SOURCE TRAIL
Evidence cited by the author
Open each link to inspect the source. Its role reflects the author’s description, not independent verification by AINewsIQ.
Shows the public model listing, architecture description, loading example, license label, and linked weight files; it does not validate the creator’s benchmark figures.
Shows the public README, model quick start and repository contents. The README says SGLang and llama.cpp backend code and setup guides will be added here.
JOIN THE CONVERSATION
Have a story worth showing your work on?
Apply to contribute original reporting or analysis. Every author and every story receives separate editorial review.
Discussion
Comments are reviewed before publication. Discuss the evidence, not the person.
No published comments yet.
Reader discussion opens with accounts. You can read the AINewsIQ story and published comments now.