DiffuSpace Open-Sources Clover-35B Diffusion Model, Outperforming Autoregressive Peers
AI lab DiffuSpace has released and open-sourced Clover-35B-A3B-Preview, the first diffusion language model with a MoE architecture. By employing iterative refinement for 'global thinking,' it outperforms the same-scale autoregressive model Qwen3.6-35B-A3B in benchmarks, particularly excelling in agentic coding and tool use while maintaining faster generation speeds.
Full Report
IT Home reported on October 29 that AI lab DiffuSpace released the first member of its Clover model family yesterday: Clover-35B-A3B-Preview, which has been open-sourced under the Apache license.
Clover-35B-A3B-Preview is a diffusion language model featuring a Mixture-of-Experts (MoE) architecture. Unlike traditional autoregressive models, it possesses "global reasoning" capabilities, refining outputs from coarse to fine through multiple iterative steps.
DiffuSpace stated that the release of Clover-35B-A3B-Preview demonstrates that diffusion language models can compete with—and even slightly outperform—autoregressive models at larger parameter scales while maintaining faster generation speeds.
The model outperforms the similarly sized autoregressive model Qwen3.6-35B-A3B across multiple benchmarks, showing particular strength in agentic coding and tool-use tasks.
References
Advertisement Disclosure: External links included in this article (including but not limited to hyperlinks, QR codes, and passcodes) are provided for informational purposes to help readers save time on selection. Results are for reference only. All IT Home articles containing external links include this disclosure.