A third of the most-liked Hugging Face Spaces no longer run. Beam reads each broken Space and writes a repair; we rebuild it and call its API. A Space only counts as revived when it returns a real result.
Every Space below crashed on Hugging Face. Each thumbnail is what the repaired app returned when we called it in our rebuild. Tap one to see the original error, Beam's diagnosis, the patch and the full output.
Type one of the 10,000 most-liked Spaces. See whether it ran when we checked, the error Hugging Face showed, the root cause Beam's triage found, and what happened if we treated it.
Most broken Spaces did nothing wrong. Spaces reinstall unpinned dependencies on every rebuild, so code that worked when it shipped now meets gradio 5 and 6, huggingface_hub 1.0, numpy 2, torch 2.6 and Python 3.13. The fix is usually small. Finding it means reading the repo, the traceback and the installed package versions together.
Beam gets all of it in one prompt: the repository with the most relevant files first, Hugging Face's error log, our rebuild log, the exact pip freeze, and the Dockerfile the Space is built from. It answers with a patch, a dated lockfile, or the verdict that the Space can't be fixed honestly. Then we check.
Beam read the error log of every broken Space in a single long-context call and named the causes, each as a regular expression. We ran its expressions over the logs ourselves, so every count is measured.
The obvious alternative needs no model at all: reinstall every dependency as it existed on the day of the Space's last commit, with uv --exclude-newer. We ran it on the same patients, with the same verifier.
Any model graded by an automatic check can learn to pass the check instead of doing the task. These are the places that could have happened in this project, and what the logs show.
Diagnosis, patch, every attempt and what the rebuilt app did, for each Space Beam treated. Failures included.