Lance is an open lakehouse file format built for multimodal AI data. It aims to replace Parquet for ML workloads with up to 100x faster random access, built-in vector indexing and data versioning โ while staying compatible with the modern data stack. The project has about 7,200 stars on GitHub and is actively developed in 2026.
AI datasets are not like analytics tables: they mix text, images, video, audio and embeddings, and training and retrieval jobs read them in random patterns that make columnar analytics formats crawl. Lance is designed for exactly that access pattern. Converting from Parquet takes a couple of lines of code, and from there you get fast point lookups and random sampling for training, a native vector index for similarity search over embeddings, and dataset versioning with time travel for reproducibility. Lance plugs into Pandas, DuckDB, Polars, PyArrow and PyTorch, works on local disks or object storage, and forms the storage layer underneath LanceDB, the embedded vector database from the same team. The format and its Rust/Python libraries are Apache-2.0 licensed, and the project positions Lance as the open table foundation for data-centric AI: one copy of your multimodal data serving analytics, training and retrieval at once.
Lance is free and open source under the Apache-2.0 license. (The related LanceDB database offers optional paid cloud services, but the Lance format itself is fully open.)
Lance is for ML engineers and data teams whose Parquet datasets have become the bottleneck โ especially teams storing embeddings and multimodal training data that needs fast random access and search.