Hugging Face relaunches Papers With Code with closed-source model support
New leaderboards now include GPT-5.5 and Mythos 5 evals—toggle them off for open-only view.
Hugging Face has relaunched paperswithcode.co as a central hub for discovering state-of-the-art results across AI domains, from 3D generation to AI agents. The platform automatically parses research papers from arXiv and Hugging Face to generate interactive leaderboards with scatter plots and tables for each benchmark—users can hover over data points to see model details.
A key new feature is support for closed-source model evaluations, which now appear alongside open models. Examples like GPT-5.5 and Mythos 5 show closed tags on their evals, and users can toggle these off to view only open-source leaderboards. The site also accepts submissions beyond arXiv, such as blog posts, enabling what Niels jokingly calls 'papers without code'.
- Automatically parses arXiv and Hugging Face papers to create benchmark leaderboards for SOTA tracking across AI fields.
- New toggle lets users hide closed-source model evals (e.g., GPT-5.5, Mythos 5) for an open-only view.
- Supports scatter plots and tables for each benchmark, with hover-over details for individual models.
Why It Matters
Brings transparency to AI benchmarks by including closed-source evaluations, while still offering a sanitized open-model view.