Clover-35B-A3B-Preview
Open-weight diffusion language models have usually trailed autoregressive ones of the same size on agentic coding; DiffuSpace claims Clover beats its own autoregressive base (Qwen3.6-35B-A3B) on SWE-bench Verified. Quick start (model card): Transformers 5.19, trust_remote_code=True, BF16 on one GPU with ~69 GB plus inference state, generate(..., block_size=32, threshold=0.95).
Model idDiffuspace/Clover-35B-A3B-Preview
PricePrice not published
Checked29 October 2026
No published training cutoff
At least333events since its release on 28 October 2026
5 historic, 65 major. 16 days of news.
DiffuSpace doesn’t publish a training cutoff, so we count from its release date. The real gap is larger.
How to call it
4 ways to use Clover-35B-A3B-Preview, with the model id each provider expects.
| Provider | Model id | Endpoint | Docs |
|---|---|---|---|
| Hugging Face (open weights, BF16 ~69 GB; custom code, trust_remote_code) | Diffuspace/Clover-35B-A3B-Preview |
huggingface.co | |
| GitHub (SGLang and llama.cpp inference backends, self-speculative decoding) | github.com | ||
| ModelScope | modelscope.cn | ||
| Web demo (application required) | chat.diffuspace.ai |
Pricing
Price not published.
No price source linked. Checked 29 October 2026.
Specs
- Input
- Text
- Output
- Text
- Context window
- 131,072 tokens
- Max output
- Not published
- Open weights
- Yes
- Licence
- Apache-2.0
- Training cutoff
- Not published
- Released
- 28 October 2026
- Status
- Preview
- Type
- Reasoning model
What stands out
What its maker marketed at launch, and what people found later, each with its source.
Block-diffusion MoE language model for agents
Text diffusion model derived from Qwen3.6-35B-A3B (35B total, ~3B active, 256 experts top-8): generates 32-token blocks in parallel by unmasking confident positions; DiffuSpace reports SWE-bench Verified 73.6, Terminal-Bench 2.1 51.7 and τ³ 65.3, above Qwen3.6-35B-A3B and the diffusion models LLaDA2.2 flash and Mercury 2.5 (vendor numbers).
Fast single-stream generation
Vendor-measured 680 tokens/s at batch size 1 in BF16 on one GB300 on internal agent workloads, and 515 tokens/s on one H200 vs 365 tokens/s for a DFlash2-accelerated Qwen of the same size.
Notes
First model from DiffuSpace, a company started by members of HKU NLP’s Dream diffusion-LM team. Preview release; thinking mode on by default; generation needs the repository’s custom code (Transformers 5.19). Long generations can repeat, so the defaults add presence, repetition and length penalties. All benchmarks are the vendor’s own; no independent leaderboard results yet.