DiffuSpace Releases Open-Source Diffusion Language Model Clover-35B-A3B-Preview

Diffusion language models are seen as a fresh route to challenge the mainstream autoregressive approach. AI lab DiffuSpace has now released Clover-35B-A3B-Preview, the first model in its Clover family, and open sourced it under the Apache license.

Unlike traditional autoregressive models, this model uses a MoE (Mixture of Experts) architecture and has a kind of "global thinking". It refines its output from vague to clear through multiple rounds of iteration, while keeping a faster generation speed.

According to the lab, the model proves that diffusion language models can compete with or even slightly beat autoregressive models at larger scales. It outperformed the same-size autoregressive model Qwen3.6-35B-A3B in several benchmarks, and is especially strong at coding agents and tool use tasks.