Clover-35B-A3B-Preview on BenchGecko · open weights diffusion MoE
DiffuSpace released Clover-35B-A3B-Preview on 2026-10-28 under Apache 2.0. BenchGecko lists the block diffusion MoE with 35B total and about 3B active parameters.
What launched
DiffuSpace, a team from the University of Hong Kong NLP lab, released Clover-35B-A3B-Preview on 2026-10-28 under Apache 2.0. It is a block diffusion language model: it drafts whole blocks of text and refines them over several passes instead of writing one token at a time.
BenchGecko added the model and DiffuSpace as a new provider today.
Specs
From the Hugging Face model card and config:
- Architecture · block diffusion, mixture of experts built on the Qwen3.6 35B A3B design
- Parameters · 35B total, about 3B active per token
- Position limit in config · 131,072 tokens
- Inputs and output · text
- Weights · open, Apache 2.0, also on ModelScope
- Serving · SGLang and llama.cpp support
Pricing on BenchGecko
There is no public API price yet and the model is not on OpenRouter. BenchGecko leaves the USD price fields empty.
Vendor-reported results
DiffuSpace reports these figures. They are self-reported and not yet independently reproduced, so BenchGecko records them as vendor claims.
- SWE-bench Verified · 73.6
- Terminal-Bench 2.1 · 51.7
- GPQA Diamond · 83.8
- Throughput · 680 tokens per second on a single GB300 card
Sources
Source: BenchGecko
Written automatically from BenchGecko's tracked data for this period; every number comes from the changelog, model, mindshare and pricing datasets. Checked against the source data before publishing.