DiffuSpace Clover-35B: An Open-Weight Diffusion LLM That Claims to Beat Qwen at Agentic Coding While Running Faster
Source: DiffuSpace's official Hugging Face blog post, "Clover-35B-A3B-Preview: A New Way for Intelligence" (published Oct 30, 2026), with links to GitHub and the Hugging Face model repo
Oct 30, 2026: Nearly every chatbot you use writes text one token at a time, left to right. DiffuSpace, a team that includes people behind the University of Hong Kong's Dream 7B diffusion model, has released a large open model that doesn't work that way. Clover-35B-A3B-Preview is a diffusion language model: it generates blocks of text in parallel and refines them, like an image model "denoising" a picture.
DiffuSpace says the bet pays off: at a similar size, the model outperforms comparable token-by-token models while generating significantly faster. The weights are open under an Apache license.
The model at a glance
Per DiffuSpace's blog:
- Architecture: Mixture-of-Experts (MoE) diffusion LLM
- Size: 35B total parameters, 3B active
- Thinking mode: supported
- License: Apache (the Hugging Face repo metadata lists apache-2.0)
- Where to get it: Hugging Face, GitHub and ModelScope
- Inference support: SGLang and llama.cpp
- Built on: DiffuSpace thanks Qwen for "its open model architectures and training ecosystem," which it says gave the research "a reliable starting point"
The headline claims (all self-reported)
Every number below comes from DiffuSpace's own testing. None has been independently verified.
Agentic coding and tool use: - 73.6 on SWE-bench Verified - 51.7 on Terminal-Bench 2.1 - 65.3 on τ³ Overall - DiffuSpace says it posted the top score on all six agent benchmarks in its comparison. That comparison covered other diffusion models (LLaDA2.2 flash, Mercury 2.5) and autoregressive Qwen3.6-35B-A3B and Qwen3.5-35B-A3B.
General capability: - 83.8 GPQA Diamond, 84.8 MMLU-Pro, 95.8 ZebraLogic, 72.0 AA-LCR
Speed: - 680 tokens/s on internal agent workloads on a single NVIDIA GB300 (BF16, batch size 1) - 515 tokens/s on a single H200, against 365 tokens/s for a Qwen model accelerated with DFlash2 - 101.8 tokens/s on MATH-500 with INT4 quantization on an RTX 5070 Ti (16 GB VRAM), against 48.5 tokens/s for similarly sized MiniCPM and Qwen models in DiffuSpace's tests
How it works (short version)
DiffuSpace says it redesigned Gated DeltaNet (GDN) for block diffusion:
- Previously generated text is compressed into a recurrent memory state.
- Tokens inside the current block attend to one another bidirectionally, so the model can look ahead within the block as well as back.
- The memory readout competes in the same softmax as the in-block tokens. That lets the model learn when to lean on earlier context and when to use the current block.
The goal is to keep long-context memory cheap while letting a whole block of tokens be "denoised" at once. That parallelism is where the speed comes from.
Why gamers and builders should care
The consumer-GPU result is the interesting one. If a 35B-class model at INT4 really runs on a 16 GB RTX 5070 Ti at around 100 tokens/s, that's a useful local coding or agent model for a gaming PC. For AI agents, DiffuSpace argues, "speed is itself a core capability": each faster step shortens the plan-act-check loop.
DiffuSpace's demos include a web agent finding driving directions, a chain of six business tasks using WebMCP tools, and a single-file HTML "desktop" with working apps. These are curated demos and shouldn't be read as benchmarks.
What's coming next
DiffuSpace says these are not released yet: - A decision model built on Clover, with a "Joint Mode" that answers up to eight related questions in one denoising pass. DiffuSpace says it will be released "at a later date." - An agent execution harness and a desktop coding client, described as "under development" - GenUI partner SigmaZ is running a closed education beta and joint post-training. The stated target is 2,000 tokens/s.
What's confirmed / not confirmed
Confirmed (DiffuSpace's official blog, Oct 30, 2026): - Clover-35B-A3B-Preview is released as an open-weight, Apache-licensed MoE diffusion LLM with 35B total and 3B active parameters. - Weights and code are on Hugging Face, GitHub and ModelScope, with SGLang and llama.cpp support. - The Clover decision model, the agent harness and the desktop client are announced but not yet released.
Not confirmed / self-reported: - All benchmark scores and speed figures are DiffuSpace's own results, several on "internal" workloads. Independent replication hasn't been published. - Real-world llama.cpp performance on consumer GPUs beyond DiffuSpace's single RTX 5070 Ti test. - The blog also asks users to "apply for access" at diffuspace.ai, alongside the open downloads. It isn't clear what the hosted access includes beyond the open weights. - It's a "Preview" release, so expect changes.
Related
Up: AI
