Reportedly: DiffuSpace releases and open-sources Clover-35B-A3B-Preview, a 35B MoE diffusion language model
Reported by Zhidongxi · not yet confirmed by the company or a second independent outlet. We update this page when it is.
DiffuSpace released and open-sourced Clover-35B-A3B-Preview, a 35B-parameter MoE diffusion language model with 3B active parameters. The company reports it ranks first among same-scale models on six agentic coding and tool-use benchmarks, surpassing Qwen3.6-35B-A3B. A decision model variant will be released later.
China context
- Original name
- 扩散智能
- Outside China
- Open weights · zhidx.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers can download the open weights from Hugging Face, GitHub, or ModelScope and evaluate the model on agentic coding tasks; the decision model's joint mode may reduce inference calls for multi-question judgment workflows.
- For investors
- DiffuSpace's open-source release and claimed benchmark leadership in agentic coding could pressure other model providers in the open-weights segment; the decision model's efficiency gains may attract enterprise adoption if independently validated.
Translated from Chinese. Quotes and facts link to the original sources.
Shenzhen-based DiffuSpace released and open-sourced Clover-35B-A3B-Preview, a diffusion language model with a Mixture-of-Experts architecture (35B total parameters, 3B active) supporting a thinking mode. The company's benchmark results show it ranks first among same-scale models on six agentic coding and tool-use tasks, including SWE-bench Verified and Terminal-Bench2.1, outperforming the autoregressive Qwen3.6-35B-A3B. DiffuSpace also introduced a decision model variant for fast, structured judgment tasks, supporting independent and joint modes; the joint mode achieved 95.0% accuracy on an internal validation set of 500 groups with 2,420 related questions, versus 89.8% for independent mode and 63.6% for Jev. In a Chinese chess match, the decision model defeated Jev. The model is available on Hugging Face, GitHub, and ModelScope, with an online experience application at chat.diffuspace.ai.
Clover-35B-A3B-Preview uses a diffusion language model (dLLM) approach with a Mixture-of-Experts architecture, activating 3B of 35B parameters. The decision model's joint mode performs global reasoning across a set of related judgments in one pass, avoiding multiple interactions and logical conflicts; on an internal validation set of 500 groups (2,420 questions), joint mode answered 95.0% correctly and got all questions right in 80.4% of groups, compared to 89.8% and 62.4% for independent mode. The model is open-sourced on Hugging Face, GitHub, and ModelScope.
Developers outside China can now access and test a diffusion-based MoE model that DiffuSpace claims outperforms a comparable autoregressive model on agentic coding benchmarks, potentially offering an alternative architecture for agent workflows. The decision model's joint mode reduces inference time for multi-question judgment tasks, which could lower costs for structured decision applications if the reported accuracy holds in independent evaluations.
DiffuSpace provides open weights for Clover-35B-A3B-Preview, enabling developers to deploy and fine-tune the model without API fees. The decision model's joint mode may reduce compute costs for batch judgment tasks by processing related questions in a single pass.
The decision model variant is slated for later release; its availability and performance on public benchmarks remain to be verified. Independent evaluation of Clover-35B-A3B-Preview on the cited benchmarks would confirm the company-reported rankings.