Ente Photos: AI search plans and self-hosting limits

Ente's 2026 community AMA covers photo management, encryption and plans for AI search. Laurence says the team has started building a small on-device LLM that would translate typed or spoken requests into search filters. His example combines a date, a location and a person. This is work in progress, while a vision language model for discussing photos remains a future idea.

For face recognition, the team describes improvements to clustering rather than replacing the underlying models. Laurence also reports that a Rust rewrite of the mobile ML stack is nearly complete but unreleased. The aim is faster processing and shared behavior across mobile, desktop and web.

Language support remains a limitation: the team acknowledges that search works better in English. Laurence proposes multilingual fine-tuning of the planned search model, provided it has some prior training in each language. AI object removal and selection are also planned rather than presented as available features.

The self-hosted discussion addresses using two storage buckets instead of three and lifting a 10 GB file size limit. Vishnu gives no delivery date and says changing replication requires additional work. Migration between instances is planned without an ETA. The team explains that encrypted content requires analysis on the user's device and warns that losing both the password and recovery key can leave it unable to recover the data.