DiffuSpace open-sources Clover-35B-A3B-Preview, a block-diffusion MoE language model that it says beats its Qwen3.6 base on agentic coding (SWE-bench Verified 73.6)
Confirmed
The takeaway
DiffuSpace, founded by members of the University of Hong Kong NLP group behind the Dream diffusion models, released Clover-35B-A3B-Preview under Apache 2.0 around Oct 28, 2026.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 2 of 5
- Last verified
- 29 October 2026
Your AI and this story
- GPT-6 Astra181 days after its cutoff
- Claude Opus 5.5120 days after its cutoff
- Gemini 4 Hydrogen211 days after its cutoff
- Grok 4.7150 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 120 days before it.
Key facts
- Architecture (model card): Qwen3.6-35B-A3B backbone kept (hybrid Gated DeltaNet / full attention, 40 layers, 256 routed experts, top-8, one shared expert); bidirectional attention inside each block; block size 32, confidence threshold 0.95
- Vendor benchmarks (blog): SWE-bench Verified 73.6, Terminal-Bench 2.1 51.7, τ³ Overall 65.3 (best of the compared models on six agent tests, against LLaDA2.2 flash, Mercury 2.5, Qwen3.6-35B-A3B and Qwen3.5-35B-A3B); GPQA Diamond 83.8, ZebraLogic 95.8, AA-LCR 72.0
- Speed (vendor): 680 tokens/s overall on internal agent data, BF16, batch size 1, one NVIDIA GB300; 515 tokens/s on one H200 vs 365 for a DFlash2-accelerated Qwen model
- Release: Hugging Face repo created Oct 24, GitHub README and benchmark image added Oct 27–28; community GGUF and MLX quantizations appeared Oct 28; Pandaily covered it Oct 29
- Lineage: Dream 7B (HKU NLP, April 2025) has 2.7M+ downloads, per DiffuSpace
What happened
A start-up formed by members of HKU NLP’s Dream team turned an open autoregressive MoE model into a diffusion model and claims it is now both better and faster at agentic coding than the model it came from. Weights, inference code for SGLang and llama.cpp, and a web demo are public.
Why it matters
Diffusion text models promise much faster generation, which matters most for agents that loop many times per task. Until now open diffusion LMs trailed autoregressive models of the same size; the claim still needs independent benchmark runs.
Sources
4 sources from 4 sites. Numbers match the chips in the text.
4 sources: 3 primary, 1 press
Primary
- DiffuSpace blog: Clover-35B-A3B-Preview (in Chinese)diffuspace.ai, official
- Hugging Face: Diffuspace/Clover-35B-A3B-Preview (model card)huggingface.co, code
- GitHub: DiffuSpace/Clover-35B-A3B-Previewgithub.com, code
Press
- Pandaily: DiffuSpace open-sources Clover-35B-A3B (Oct 29)pandaily.com, press
Changes
- Filed (model card, blog and GitHub read directly; news agent)