DiffuSpace releases Clover-35B-A3B-Preview diffusion language model
Original title: Diffuspace/Clover-35B-A3B-Preview — new model trending #30 on Hugging Face
The text-only model uses about 3B active parameters per token and a 128K context window.
DiffuSpace's first release, Clover-35B-A3B-Preview, is an open diffusion language model derived from Qwen3.6-35B-A3B. It pairs a 35B mixture-of-experts backbone with block diffusion, LowConfidence decoding, and a hybrid Gated DeltaNet and full-attention design of 40 layers and 256 routed experts. Weights are on Hugging Face, with inference code for Transformers, SGLang, and llama.cpp. The company says it led open diffusion models on benchmarks in its launch blog.
Unconfirmed: Leading open-source diffusion-model performance is DiffuSpace's own reported benchmark claim.
- Defaults are block size 32 and a confidence threshold of 0.95.
- BF16 weights need about 69 GB plus inference state.
- The example install uses Transformers 5.19.0 and PyTorch 2.10 or later.
- Context is capped at 131,072 tokens for prompt and response together.
- Vendors
- DiffuSpace
From the original article
4 min read at huggingface.cotext-generation · 1,030 downloads · 139 likes · license apache-2. 0
By Diffuspace · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.