Papers with Code From Meta's 2025 Shutdown to Hugging Face's Apache 2.0 Rebuild

As of October 28, 2026, the machine learning catalog and benchmark tracker Papers with Code operates as an open-source web application maintained by Hugging Face at paperswithcode.co, with its full application codebase published at github.com/huggingface/paperswithcode on October 26, 2026. The repository documentation states that the project "recreates the original paperswithcode.com, which Meta retired in 2025, and keeps it up to date with automated imports" from Hugging Face, arXiv, GitHub, and public leaderboards.

Understanding the current state of the platform requires separating three distinct assets that emerged after Meta ended the legacy service in 2025. The live site at paperswithcode.co is a clean-room software rebuild created at Hugging Face rather than a release of Meta's proprietary legacy codebase. Alongside that new application sit the frozen historical dataset dumps in paperswithcode/paperswithcode-data and the pwc-archive organization on the Hugging Face Hub, which carry different provenance and licensing rules from the Apache 2.0 software repository opened in October 2026.

How the 2025 Shutdown Split Redirects From Historical Data Snapshots

When Meta retired the original paperswithcode.com platform in mid-2025, visitors attempting to load the root domain were routed via an HTTP 302 redirect to huggingface.co/papers/trending. Community members documented the change on August 14, 2025, in Issue #116 of the paperswithcode-data repository, noting that the Trending Papers feed lacked the task hierarchies, method taxonomies, and historical evaluation tables of the retired website.

Hugging Face co-founder and CTO Julien Chaumond (julien-c) addressed the discussion in that thread on September 4, 2025, explaining that "we did not pretend we had built a 1:1 replacement to Papers With Code" when proposing the domain redirect to Meta instead of letting the domain expire. Between September 10, 2025, and September 14, 2025, the pwc-archive organization on the Hugging Face Hub preserved the final public data dumps from paperswithcode/paperswithcode-data, covering papers-with-abstracts, links-between-paper-and-code, evaluation-tables, methods, and datasets. On September 12, 2025, Hugging Face also initialized the huggingface/paperswithcode repository on GitHub (2025-09-12T20:59:15Z), laying the groundwork for a dedicated replacement while Chaumond deployed URL redirects on December 12, 2025, and December 16, 2025, to route legacy paperswithcode.com/paper/ and paperswithcode.com/dataset/ links to arXiv-indexed Hugging Face pages.

Even with those December 2025 URL routing rules in place, paperswithcode.com itself remained a redirect endpoint rather than an active leaderboard database. As of October 28, 2026, requests to the root paperswithcode.com domain still return an HTTP 302 redirect to huggingface.co/papers/trending, while the reconstructed catalog and benchmark tracker run independently on paperswithcode.co and the secondary short domain paperswithco.de.

Inside the May and June 2026 Relaunch on paperswithcode.co

The dedicated software rebuild went live on May 17, 2026, led by Hugging Face machine learning engineer Niels Rogge (nielsr). In his May 24, 2026 relaunch post on the Hugging Face Blog, Rogge wrote: "It's been one week since I launched paperswithcode.co, a revival of the website we all loved." That initial May 2026 release restored multi-metric benchmark tables—such as tracking both Word Error Rate (WER) and Inverse Real-Time Factor (RTFx) on the Open ASR Leaderboard, or frames per second (FPS) alongside mean average precision (mAP) on COCO object detection—while adding support for non-arXiv external papers submitted through paperswithcode.co/submit and banner indicators for paper lineage.

On June 22, 2026, Rogge detailed a second wave of structural additions in a follow-up update on the Hugging Face Blog, introducing the short domain paperswithco.de and restoring State-of-the-Art (SOTA) badges for papers ranking within the top three scores of a benchmark. The June 22, 2026 update also revised how research rankings are ordered across the catalog by combining GitHub star velocity with the activity score of linked Hugging Face models, datasets, and Spaces, replacing the older formula that relied solely on GitHub star velocity.

Beyond paper-reported tables, the June 22, 2026 release introduced third-party external evaluations so benchmark pages can display independent evaluation runs alongside numbers reported by paper authors. That separation between self-reported metrics and external verification parallels broader shifts in AI benchmarking and leaderboard governance as well as automated tools entering scientific peer review workflows. To populate historical tasks and evaluation tables, Rogge noted on June 22, 2026, that legacy records from pwc-archive are imported gradually into the new catalog alongside scheduled syncs from active leaderboards.

Clean-Room Application Stack Versus Legacy Dataset Archives

On October 20, 2026, Rogge merged Pull Request #609 to rewrite the repository documentation for public release and Pull Request #611 to place the huggingface/paperswithcode codebase under the Apache License 2.0. Six days later, on October 26, 2026 (2026-10-26T10:04:52Z), Pull Request #722 linked the public GitHub repository directly from the paperswithcode.co navigation bar, completing the transition of the web application into an open-source project.

Architecturally, the repository opened on October 26, 2026, is a self-contained web stack built on Python 3.12, FastAPI, SQLAlchemy 2, and Pydantic on the backend, paired with a React 18, TypeScript, Vite, TanStack Query, shadcn/ui, and Tailwind CSS single-page frontend. Search and retrieval run on PostgreSQL 16 using pg_trgm for lexical matching and pgvector for semantic embeddings. A maintenance command-line interface (pwc) and automated workflow jobs handle daily paper ingestion, OpenAlex organization tagging, and leaderboard score validation for open-weights model releases and benchmark tracking. Production hosting relies on immutable nginx releases backed by four private Hugging Face storage buckets (huggingface/pwc-papers-content, huggingface/pwc-thumbnail, huggingface/pwc-paper-embeddings, and huggingface/paperswithcode-backups).

That software architecture is legally and operationally distinct from the legacy dataset archive. The LICENSE file in huggingface/paperswithcode applies the Apache License 2.0 to the new application's source code, documentation source, and configuration files, whereas the repository documentation (README.md and docs/data-workflows.md, updated October 20, 2026) only references license as an optional metadata field on individual dataset entries (DatasetCreateRequest). By contrast, the historical dumps preserved in paperswithcode/paperswithcode-data and mirrored at huggingface.co/pwc-archive—where the organization card notes that "The data in this repo corresponds to the last publicly available snapshot of the PWC datasets"—remain governed by the Creative Commons Attribution-ShareAlike 4.0 license (CC-BY-SA 4.0) specified in the paperswithcode-data repository.

Public Catalog API Limits and Local Deployment Architecture

Developers cloning huggingface/paperswithcode for local execution initialize an empty PostgreSQL 16 schema with vector and pg_trgm extensions rather than downloading a pre-populated database dump, using the pwc maintenance CLI or the standalone read-only pwc-cli package to index and query records. For programmatic read access against the hosted paperswithcode.co instance, the platform exposes a REST API under /api/v1/ that maintains compatibility with the original paperswithcode-client library.

In the hosted production environment documented in docs/public-api.md as of October 28, 2026, the interactive public API reference is served by the frontend at paperswithcode.co/docs/api. The OpenAPI schema is generated at /api/v1/openapi.json, while the bare /docs path returns HTTP 404 in production. Nginx traffic controls on paperswithcode.co enforce per-IP rate limits of 120 requests per minute (with a 60-request burst allowance) for catalog reads across /api/v1/papers, /api/v1/tasks, /api/v1/datasets, and /api/v1/evaluations, alongside 60 requests per minute (with a 20-request burst allowance) for list and search queries requiring mode=keyword, returning a two-second Retry-After header when limits are exceeded.