clover 35b: Open Diffusion LLM Claims 680 Tokens/Sec

The clover 35b release is the loudest diffusion LLM drop in a while. DiffuSpace published Clover-35B-A3B-Preview on October 30: an open-weight Mixture-of-Experts diffusion language model with 35 billion total parameters and about 3 billion active, Apache licensed, and pitched as faster than similar-size autoregressive models while matching or beating them on agent benchmarks.

What actually happened

DiffuSpace introduced Clover as the first version of its model family in a Hugging Face blog post. The team says it grew out of the group behind Dream 7B, a diffusion language model from the University of Hong Kong's NLP group that has passed 2.7 million downloads.

The question the clover 35b launch tries to answer is simple: can a diffusion LLM stay competitive at scale while keeping its speed edge? DiffuSpace says yes. Instead of generating one token at a time left to right, Clover generates blocks of text in parallel and refines them, which is where the speed claim comes from.

Benchmark numbers in the post include 83.8 on GPQA Diamond, 95.8 on ZebraLogic, 72.0 on AA-LCR and 65.3 on a tau-cubed overall score. The team says Clover posted the highest scores among the compared models on all six agent benchmarks, beating fellow diffusion models LLaDA2.2 flash and Mercury 2.5 and the autoregressive Qwen3.6-35B-A3B and Qwen3.5-35B-A3B.

How clover 35b works

The architectural trick is a redesigned Gated DeltaNet for block diffusion. Earlier context gets compressed into a recurrent memory state, while tokens inside the current block talk to each other through bidirectional attention. The model learns when to lean on history and when to focus on the block it is denoising.

The model card adds practical details: roughly 3B active parameters per token, a 128K context window, and support for Transformers, SGLang and llama.cpp. A separate GGUF repo tags self-speculative decoding and lists quantized files of roughly 23 GB for Q4_K_M and 37.9 GB for Q8_0, so the full-precision weights need far more memory than a typical gaming GPU.

Diffusion LLMs used to be the cool research demo. Clover is the first one in a while that people can actually download and run at this size.

Why the hype is real, and where it stops

Open weights, an Apache license and an agentic-coding pitch is exactly the combo that gets shilled on X. Speed matters for agents because they call the model many times per task, so a fast model at a similar quality level can change the cost math.

Still, every number here is self-reported. DiffuSpace measured speed on internal agent workloads with specific hardware, and the benchmarks were chosen by the team. Independent evals will tell the real story. We saw the same pattern with recent open releases like nanbeige4.3 33b and liquid ai lfm3: early claims, then community testing.

It is also labeled Preview for a reason. Expect rough edges in tooling and quantized builds.

Who should care

Developers running local or self-hosted models are the obvious audience. A 3B-active MoE that claims frontier-ish agent scores is the kind of release that ends up inside coding assistants, internal tools and hobby agent stacks within days.

Investors in AI-adjacent tokens should care less than the timeline suggests. A model release is not a token, and open weights mean nobody owns the upside in a tradeable way.

What to do as a reader

If you build agents, try clover 35b on your own tasks before swapping anything in production. Compare latency and accuracy against your current model on the same prompts, same hardware.

If you are on consumer hardware, start with the GGUF Q4 build and llama.cpp rather than the full BF16 weights. Check the GitHub repo for the exact launch flags.

If you just follow the space, watch for third-party benchmark runs over the next week. That is when we will know whether the speed and agent scores hold up outside DiffuSpace's own lab.

And ignore any "Clover token" that pops up on launchpads. DiffuSpace has not announced one, and AI model names get hijacked by memecoins within hours.