The relevant leaderboard is the one here: Open LLM Leaderboard - a Hugging Face Space by HuggingFaceH4.

For a model to be considered a "diffusion-based" model it should be explicitly trained with a diffusion-based training objective (e.g., denoising score matching, etc.) and sampling should happen with a diffusion-based algorithm (e.g., multi-step sampling with SDE solvers or Predictor-Corrector methods, or single-step consistency-based or distilled sampling, etc.).

Some recent (non-exhaustive) examples of diffusion-based LLMs are surveyed here: [2303.07576] Diffusion Models in NLP: A Survey (arxiv.org).

10/30/2023 Update: There are still no diffusion-based language models in the top-10 leaderboard, but there has been marked progress, with some models outperforming autoregressive models for code generation, e.g. as in [2310.17680] CodeFusion: A Pre-trained Diffusion Model for Code Generation (arxiv.org).

11/9/2023 Update: After some controversy, the above-mentioned paper ended up being retracted from arXiv. Diffusion for language still has a good way to go.

2/16/2020 Update: No language diffusion models have made it up to top-10 so far. The best language diffusion models so far are still at GPT-2 scale.