DiffuSpace open-sources Clover-35B-A3B-Preview diffusion language model

Summary

DiffuSpace has open-sourced Clover-35B-A3B-Preview, a MoE-based diffusion language model, under the Apache license. The company claims that the model outperforms the Qwen3.6-35B-A3B autoregressive model of the same scale in many benchmarks and produces faster outputs.

Why it matters

Diffusion-based language models have historically lagged behind autoregressive architectures in large-scale applications. A reliable open-source version that matches or exceeds the performance of its 35B-A3B autoregressive counterpart would validate an alternative production paradigm. The Apache license facilitates non-autoregressive experimentation with a competitive parameter budget for researchers and practitioners.

Highlights

• An MoE diffusion language model released under the Apache license, marking the first member of the Clover family
• Claims superiority over Qwen3.6-35B-A3B, particularly in agentic coding and tool usage
• Positions 'global thinking' via 'multi-turn iterative refinement' as a speed and quality advantage over token-by-token autoregression

Critical take

The article is essentially a vendor press release, lacking independent benchmark tables and ablation data on the wall-clock cost of 'multi-turn iteration' compared to a single autoregressive pass. The 'Preview' label indicates that the weights are not yet final. The self-selected comparison against only one Qwen variant, without mentioning other 35B-class models or inference hardware, makes it difficult to independently verify the 'faster and better' claim from this source alone.