First open-weight production dLLM: Clover-35B hits 680 tps on one GPU, beats same-size AR models on SWE-bench
taoyds · x · 2026-10-28
DiffuSpace released Clover-35B-A3B-Preview, billed as the first open-weight production diffusion language model (dLLM): 35B total params (3B active), 680 tokens/s at BF16 on a single card, available on Hugging Face and ModelScope.
- Strong agent/coding results: SWE-bench Verified 73.6, Terminal-Bench 2.1 51.7, Terminal-Hard 38.1, τ³ Overall 65.3 — well above LLaDA2.2 flash, Mercury 2.5, and similar-size Qwen AR models
- Knowledge/reasoning: GPQA Diamond 83.8, MMLU-Pro 84.8, ZebraLogic 95.8
- Weak spot: HLE text-only 12.8, behind Qwen
Diffusion LLMs are starting to genuinely compete with autoregressive models on agentic coding workloads.
More from Models
- OpenAI $200-per-month Pro users hitting "model at capacity, try a different model" errors — johnseach · 2026-10-28
- Evil 27B Too Risky to Run Autonomous? User Fears Wiped Drive Over a Game — teortaxesTex · 2026-10-28
- DeepSeek V5 rumor debunked: code strings were just September assert tests — teortaxesTex · 2026-10-28
- ChatGPT Pro usage limits being halved from 20x to 10x, $200 subscribers grumble — kimmonismus · 2026-10-28
- Abliterated models terrify me, but so do safety filters: paraschopra on AI's fairness asymmetry — paraschopra · 2026-10-28
- New Paper Shows Gemini-Found Softmax Bug Fix Still Fails Property Tests — brandondamos · 2026-10-28