Papers with Code, Which Meta Shut Down, Has Been Open-Sourced by Hugging Face

Papers with Code is now open source. Under the Apache 2.0 license, the repository lives under Hugging Face’s organization, and Niels Rogge announced it on X. The interesting part isn’t just that the code is open—Meta shut down the original site in 2025, and Hugging Face rebuilt it by automatically importing arXiv, GitHub, HF Hub, and public leaderboards. The live site is already running at paperswithcode.co.

For those unfamiliar with this tool: Papers with Code maps papers, code, models, datasets, and leaderboards into one place. Whether a paper has code, has run benchmarks, and what the results are can be checked at a glance, without jumping back and forth between arXiv and GitHub.

The repository has quite a lot to offer:

Paper catalog. arXiv and external papers, linked to official code, project pages, HF models, and datasets, with thumbnails, one-sentence summaries, and a Markdown reader.
Taxonomy. Tasks, methods, datasets, frameworks, conferences, institutions, and authors can all be browsed, with trend views.
Leaderboards. Paper-level evaluation results, with external leaderboards synchronized on a schedule.
Hybrid search. PostgreSQL’s pg_trgm for lexical retrieval and pgvector for semantic retrieval.
Public API. The path is /api/v1/, compatible with the original Papers with Code API clients; OpenAPI docs are at /docs.
Optional research chat. Based on catalog content, grounded through the Papers with Code MCP tool.
Automated curation. Scheduled jobs import papers daily, link tasks and methods, refresh stars and citations, and propose new evaluations.

The tech stack isn’t fancy: a React/TypeScript frontend, a Python 3.12 + FastAPI backend, and PostgreSQL 16 for data storage. Maintenance uses the uv-managed pwc CLI; for example, uv run pwc paper index 1706.03762 indexes a paper. Setting up the dev environment isn’t complicated either: clone, uv sync, create the database, initialize the schema, and run one command each for frontend and backend. Deployment uses nginx plus immutable releases; in production, article content, thumbnails, and database backups are stored in Hugging Face’s own bucket.

Someone online pointed out a key judgment: the license is the real news. Previously people could only scrape the site; now they can legally mirror leaderboards and paper-code mappings, without worrying that some site might disappear again one day. Others noticed a more practical issue: dead links can be fixed—after the original site shut down, many people hit broken paper links and could only close the tab. Still others are asking governance questions: who merges disputed leaderboard entries, and by what standard, doesn’t seem to be clearly documented yet. For people outside the field, the value is more direct: when a tool claims to be powerful, looking up its papers, code, and evaluation results on Papers with Code is a concrete way to verify its real level.

Open-sourcing is one thing; having someone continuously maintain it is another. Hugging Face now does both. Meta can retire a product, but once the community takes over, the fate of this kind of infrastructure no longer depends on one company’s roadmap adjustments.

Repository: https://github.com/huggingface/paperswithcode
Live site: https://paperswithcode.co