Clover-35B-A3B: DiffuSpace releases open-source 35-billion-parameter diffusion language model

DiffuSpace released Clover-35B-A3B-Preview this Wednesday, October 28, at around 4:42 a.m. Paris time, an open language model that doesn’t work like most others. Instead of writing text one token at a time (what’s called an “autoregressive” model), it generates it through diffusion: it starts from a draft of entire blocks and refines it over several passes, in parallel. The team, from the NLP lab at the University of Hong Kong, had already released Dream 7B in 2025.

The weights are published under the Apache 2.0 license on Hugging Face and ModelScope, with the code on GitHub.

What we know about the model

• Architecture: mixture of experts (MoE), 35 billion parameters total, of which about 3 billion active per token, making it lightweight to run for its size.
• Base: it builds on the architecture of Qwen3.6-35B-A3B from Alibaba, adapting its memory mechanism (Gated DeltaNet) so it reads context in both directions within each block.
• Usage: reasoning mode, supported by SGLang and llama.cpp, two common tools for running models locally or on a server. Online access is offered on request.

The promises, to be verified

The main selling point is speed. DiffuSpace claims 680 tokens per second on a single NVIDIA GB300 card (in BF16, one request at a time, on its internal agent data), and 515 tokens/s versus 365 for a same-size Qwen model accelerated with the DFlash2 technique on an H200 card. In compressed form (INT4), the model would run at 101.8 tokens/s on a consumer-grade 16 GB RTX 5070 Ti.

On quality, the developer claims 73.6 on SWE-bench Verified and 51.7 on Terminal-Bench 2.1, two coding and agent benchmarks, as well as 83.8 on GPQA Diamond. It says it outperforms Qwen3.6-35B-A3B on several tests. These figures all come from DiffuSpace: no independent evaluator has reproduced them yet.

The post also presents a decision model derived from Clover, designed to chain several related judgments in a single pass, along with a future agent environment and a coding application, both announced as “in development.”