DiffuSpace launches Clover-35B-A3B-Preview diffusion language model with enhanced performance and speed

On October 28, 2026, the DiffuSpace AI lab unveiled its first model in the Clover series, Clover-35B-A3B-Preview, and made it available under the Apache License. This diffusion language model employs a MoE architecture, allowing it to process tasks with a 'global thinking' approach, generating results from vague to specific through multiple iterations. DiffuSpace claims that Clover-35B-A3B-Preview outperforms self-regressive models like Qwen3.6-35B-A3B in benchmark tests, particularly in agent programming and tool usage tasks, while maintaining faster generation speeds. The model's release marks a significant step in diffusion models' ability to compete with and surpass self-regressive models in scale and efficiency.