DiffuSpace Releases Clover Open Diffusion Model

DiffuSpace has released Clover-35B-A3B-Preview, an open-weight diffusion language model that it says beats a similarly sized Qwen model on agent and coding tests while writing text much faster. The weights are on Hugging Face under an Apache licence, with code on GitHub, according to the company’s launch blog.

The team has a track record in this niche. DiffuSpace says that in April 2025 the NLP Group at The University of Hong Kong released Dream 7B, a diffusion model whose family has since passed 2.7 million downloads on Hugging Face, and that “members of the team began a new chapter at DiffuSpace”. Clover is the first version of what it calls DiffuSpace Clover.

What diffusion changes

Most chatbots are autoregressive: they write one token, a word or part of a word, at a time. Diffusion models instead start with a rough block of text and refine many tokens in parallel, which is why their backers pitch them as faster. The question DiffuSpace sets itself is whether that speed edge survives when the model is scaled up. Its answer, unsurprisingly, is yes.

Clover is a mixture-of-experts model with 35 billion parameters in total but only about 3 billion active for each token, so it runs far lighter than its size suggests. It supports a “thinking” mode. The Hugging Face card says the model is derived from Alibaba’s Qwen3.6-35B-A3B, is text-only, handles up to 131,072 tokens of context and weighs about 69GB in BF16 format. DiffuSpace thanks Qwen in its blog for “a reliable starting point for the research”.

Support for SGLang and llama.cpp, two widely used open-source inference engines, is included at launch. In a post on X, DiffuSpace called Clover “the first open-weight production diffusion language model (dLLM) with 35B total parameters”.

The company’s numbers

All of the following scores come from DiffuSpace’s own tests. It says Clover posts the highest score among the models it compared on all six agent benchmarks it ran, including 73.6 on SWE-bench Verified, 51.7 on Terminal-Bench 2.1 and 65.3 on τ³ Overall. The comparison set includes two other diffusion models, LLaDA2.2 flash and Inception’s Mercury 2.5, as well as Qwen3.6-35B-A3B and Qwen3.5-35B-A3B. On general tests it reports 83.8 on GPQA Diamond, 95.8 on ZebraLogic and 72.0 on AA-LCR, a long-context test.

On speed, DiffuSpace says Clover generated 680 tokens a second on internal agent workloads on a single NVIDIA GB300 GPU, at BF16 precision with one request at a time. On a single H200 it reports 515 tokens a second, against 365 for a Qwen model sped up with a technique called DFlash2.

Against smaller on-device models, including MiniCPM-V 4.7, LFM2.5, Nemotron 3.5 Lightning and Gemma 4, it says Clover topped six of nine benchmarks. A version quantised to INT4 ran at 101.8 tokens a second on a 16GB RTX 5070 Ti graphics card, the company says, against 48.5 in its comparison tests with similarly sized MiniCPM and Qwen models.

DiffuSpace also says a separate “decision model” will be released later, and that an agent harness and a desktop coding client “will be rolled out shortly”.

Built on Qwen, measured against last season’s rivals

Clover is the latest sign that diffusion is moving from research curiosity to usable product. Inception launched Mercury 3 on 19 October with a speed pitch of its own, and Stanford’s Hazy Research lab has shown text written in a few parallel steps. What sets Clover apart is that anyone can download and run it, joining a crowded field of small open models such as Liquid AI’s LFM3 family.

Two caveats temper the claims. First, Clover is not built from scratch. It starts from Qwen3.6-35B-A3B and turns it into a diffusion model, so its wins over Qwen partly measure how well DiffuSpace converted Qwen’s own knowledge. That is a legitimate and cheaper route, but it also means Clover inherits whatever is in its base model, a point that matters to buyers who have questioned Qwen’s behaviour.

Second, the comparison set is already dated. DiffuSpace tests against Mercury 2.5 rather than Inception’s newer Mercury 3, and every score is self-reported on settings it chose, including its speed tests on internal workloads. Independent runs on public leaderboards would settle whether the agent scores hold up. The practical test is simpler: with weights and llama.cpp support out on day one, developers can now check the 680 tokens a second for themselves.