Meta shuts down Papers with Code

Papers with Code was a website that joined three things researchers care about: papers, the code that implements them, and leaderboards that rank methods on standard benchmarks. When it went offline, the site carried 9,327 benchmark leaderboards, 79,817 links between papers and code, and 5,628 datasets, all served from a single canonical URL.^[1]

That combination is what made the site work as an index. A number on a leaderboard is only checkable if a reader can move from the score to the paper that produced it and then to the code that reproduces it; when those pieces sit in separate places, verification becomes slower and often simply does not happen.

How the shutdown happened

On 24–25 July 2025 Meta sunsetted the site without prior notice. Users who tried to reach it were met with Bad Gateway 502 errors, and GitHub issues filed about the outage went unanswered.^[1]

Within days paperswithcode.com began redirecting to Hugging Face's Trending Papers feature. The 9,327 benchmark leaderboards, 79,817 paper-to-code linkages and 5,628 datasets were no longer served from the canonical URL.^[1]

On 25 July 2025 Hugging Face co-founder and CTO Julien Chaumond announced a new partnership with Meta to replace Papers with Code with Trending Papers.^[3]

What was rescued

Community members acted quickly. The final public data snapshots were preserved as JSON dumps on GitHub, in the paperswithcode/paperswithcode-data repository, and as the pwc-archive organization on the Hugging Face Hub, which describes itself as the last publicly available snapshot of the PWC datasets.^[2] Historical data also survived as 2021-era imports into the Open Research Knowledge Graph.^[1]

Keeping copies in two places, a code repository and a model-and-dataset hub, means the records do not depend on a single company keeping a server running. A snapshot is nonetheless a fixed copy: it holds what the site contained at one moment, and it does not keep the leaderboards current as new results appear.

Reaction and consequences

The shutdown drew a sharp reaction from researchers who relied on the leaderboards to track progress and to check reproducibility. No successor replicates the unified paper–code–benchmark–dataset linkage.^[1]

That gap is the substance of the event. The papers remain published, the code remains in repositories and the benchmarks keep running; what stopped being served from the canonical address was the layer that joined them together and made them findable from one place. The replacement, Trending Papers, is a Hugging Face feature rather than a reconstruction of the old site, and the archived snapshots — the GitHub JSON dumps, the pwc-archive organization and the earlier imports into the Open Research Knowledge Graph — are what remain of the original record.^[1][2]