The artificial intelligence laboratory DiffuSpace has introduced its latest development, the Clover-35B-A3B-Preview model. This release marks the debut of the company's new Clover series, which is now available under the Apache open-source license.
The model utilizes a Mixture-of-Experts (MoE) architecture and functions as a diffusion language model. Unlike standard autoregressive models, it employs a global reasoning approach, refining outputs through iterative processes that move from general concepts to specific details.
According to DiffuSpace, the model demonstrates that diffusion-based architectures can compete with or outperform traditional autoregressive models at larger parameter scales. It also maintains higher generation speeds during operation.
In performance evaluations, Clover-35B-A3B-Preview has surpassed the Qwen3.6-35B-A3B model of a similar scale. The developers noted that the new model shows particular proficiency in tasks related to agent-based programming and the utilization of external tools.
