Shenzhen's DiffuSpace Open-Sources Clover-35B Model, Claims Lead Over Autoregressive Peers in Agent Benchmarks
DiffuSpace has released and open-sourced Clover-35B, a MoE diffusion language model. Vendor tests show it outperforms same-size autoregressive models like Qwen3.6 in SWE-bench and other agent tasks.
On October 28, 2026, Shenzhen-based AI firm DiffuSpace announced the release and open-sourcing of its latest diffusion language model, Clover-35B-A3B-Preview. The model utilizes a Mixture-of-Experts (MoE) architecture with 35 billion total parameters and 3 billion active parameters, supporting a thinking mode.
According to benchmark results published by the company, Clover-35B-A3B-Preview ranked first among same-level models in six agentic coding and tool-use tasks, including SWE-bench Verified and Terminal-Bench2.1. Its performance reportedly exceeds that of other diffusion models like LLaDA2.2 flash and same-sized autoregressive models such as Qwen3.6-35B-A3B. Specifically, it achieved a score of 73.6 on SWE-bench Verified.
In terms of inference efficiency, internal tests showed the model reaching a generation speed of 680 tokens/s on a single NVIDIA GB300 GPU at BF16 precision. Across 49 agent tasks, the median execution time was reported as 1.52 seconds. Additionally, DiffuSpace previewed a decision-making model for structured judgment tasks, claiming a 95.0% accuracy rate for its 'joint mode' in internal validation.
All performance metrics are self-reported by the vendor and have not yet been independently reproduced. The model weights are available on Hugging Face and GitHub.