Indexing a year of video locally on a 2021 MacBook with Gemma4-31B (50GB swap)
A recent project detailed on Hacker News involved indexing a year of personal video footage, comprising over 1000 clips and 1TB of data, entirely on a 2021 MacBook Pro M1 Max. The developer leveraged a local Gemma4-31B model, quantized to 4-bit, which operated using 20GB of RAM and an additional 50GB of swap space. The process entailed extracting key frames from each video, feeding them to the local LLM to generate descriptive captions and embeddings, and then storing this metadata in a local Qdrant vector database. This comprehensive local setup allowed for the creation of a fully searchable personal video archive, highlighting the potential for on-device AI to manage and organize large personal media collections without cloud dependency.
Developers can now explore practical applications of large language models on consumer hardware for local data processing and personal media management.