DiffuSpace has published Clover-35B-A3B-Preview, its first model, presenting it as the most advanced open-source diffusion language model at release, at least on the benchmarks listed in its announcement.
Unlike most language models, which produce text one token at a time, Clover generates whole blocks of 32 tokens. The system starts from masked positions and fills in parallel those it is most confident about, then moves to the next block. The claimed result is faster generation, though the model is still a preview and performance needs independent verification.
Image: Diffuspace — https://huggingface.co/Diffuspace/Clover-35B-A3B-Preview
Underneath is a mixture-of-experts architecture: 35 billion total parameters, with about 3 billion active per token. According to the model card, the structure has 40 layers, 256 routed experts with top-8 routing per token, and one shared expert per layer. Context reaches 131,072 tokens across prompt and response. These are figures reported by the source and not yet independently verified.
Clover derives from Alibaba's Qwen3.6-35B-A3B. DiffuSpace adds block diffusion inference, bidirectional GDN parameters and generation controls. To limit repetition in very long responses, the model includes repetition penalties and gradual biases toward ending the response and the thinking phase.
Running it locally requires about 69 GB for the BF16 weights, so a high-end GPU. DiffuSpace recommends its SGLang implementation for serving and evaluation, since the standalone PyTorch path can produce slightly different text under the same settings.
The upstream weights are licensed under Apache 2.0, Alibaba's license, which DiffuSpace keeps in the repository together with attribution notices.
Image: Diffuspace — https://huggingface.co/Diffuspace/Clover-35B-A3B-Preview
The release fits a wave of diffusion models: Inception's Mercury 3 and Amazon's ALoDLM-8B also rely on parallel generation. The Qwen3.6 base is already in Alibaba's catalogue, and Qwen3.8-27B shows the open-weight branch of the family still growing.
DiffuSpace also announces a decision-model variant to follow later. No dates or details are available yet.
No comments yet. Be the first!