Liquid AI has developed a local privacy filter: wipes out 40 types of sensitive information in one go in 16 languages

According to 动察 Beating monitoring, Liquid AI has launched a privacy filtering model with only 350 million parameters. It can scan a full paragraph at once, identify 40 types of sensitive information such as names, phone numbers, ID numbers, bank cards, API keys, etc., support 16 languages, and can run directly on a local device.

Liquid AI compared it against models like OpenAI Privacy Filter in 6 sets of tests. According to official results, it outperformed the OpenAI Privacy Filter in 5 sets and was only behind in one set.

The model's weights have been released and can be downloaded directly from Hugging Face. Users can first scrub sensitive information locally, then send the content to cloud AIs like ChatGPT and Claude without uploading the original private data.