Most language models write one token after another. Clover is a block diffusion model: it starts each block of 32 tokens as masked positions and fills in the ones it's confident about (above a 0.95 threshold) in parallel, then moves on to the next block. It's derived from Qwen3.6-35B-A3B, so it keeps that model's mixture-of-experts layout: 35B total parameters, about 3B active per token, and a 128K context. It's text-only.
DiffuSpace's launch post claims it beats the base Qwen model and other diffusion models (LLaDA2.2 flash, Mercury 2.5) on six agent benchmarks, including 73.6 on SWE-bench Verified and 51.7 on Terminal-Bench 2.1. On speed, it reports 680 tokens/s on one GB300 and 515 tokens/s on one H200 on internal agent data, against 365 tokens/s for a Qwen setup with DFlash2.
Weights are on Hugging Face, GitHub and ModelScope under Apache 2.0. The model card's quick start uses Transformers 5.19.0 with trust_remote_code and needs about 69 GB of GPU memory for the BF16 weights. DiffuSpace recommends its own SGLang build for serving, and there's a llama.cpp path for local runs.