llm A field note by Vikrant Sharma
Someone is archiving deleted LLM models before they vanish
Pirate Face archives open-weight language models that companies try to delete. Turns out model takedowns happen more often than I thought.
A service called Pirate Face is archiving open-weight language models that get deleted from official repositories. Companies release models, then pull them down for compliance reasons, licensing disputes, or quiet pivots. Pirate Face mirrors them before they disappear. This is not about pirating commercial APIs. These are models released under permissive licenses, then yanked. Meta published Llama weights openly, then restricted distribution. Stability AI released models, then removed old versions. Pirate Face treats model weights like the Internet Archive treats websites. The practical angle: if you fine-tuned a model that later gets pulled, your deployment breaks unless you cached the checkpoint. Reproducibility dies when the base model vanishes. Academic papers cite model versions that no longer exist. Pirate Face keeps a receipt. The legal angle is messier. Hosting a deleted model might be technically allowed under the original license, but companies delete models to cut liability. If a model was trained on contested data, archiving it preserves the legal exposure. Pirate Face is betting that preservation beats caution. I checked the archive. It has early Llama versions, some Stability models, a few Chinese LLMs that disappeared after export control talk. The site does not explain who runs it or where the servers are. That opacity is probably intentional. This reminds me of arXiv for model weights. Researchers assume published models stay published, but companies are not academic institutions. They pull models when convenient. If you depend on a specific checkpoint for inference or research, you now cache it locally or trust a third-party mirror. Pirate Face is that mirror, run by someone who thinks model weights should not vanish quietly.
