LanceDB Multimodal Lakehouse

LanceDB Inc. develops a multimodal lakehouse for AI built on the open-source Lance columnar format. It targets teams that must manage images, video, text, embeddings and sensor data for search, feature engineering and model training, which traditional data lakes and standalone vector databases handle poorly.

LanceDB stores raw multimodal data, embeddings and features in one table on object storage, with vector, full-text and SQL retrieval, Git-style branching and RaBitQ quantization. It reaches 100K+ queries per second, tables of over 100 billion rows and search at 10-billion-vector scale. It integrates with DuckDB, Spark, Ray, Hugging Face Hub and LangChain, and is used by Netflix, Uber, Runway, Midjourney, ByteDance Volcano Engine, WeRide and UBS.

• Lance format: open-source columnar format optimised for multimodal data and random access.
• Vector and hybrid search: vector, full-text and SQL retrieval with multi-bit RaBitQ quantization.
• Feature engineering: compute and store features next to raw data at petabyte scale.
• Training data loading: stream data to PyTorch and LeRobot training jobs.
• Git-style branching: branch and version datasets for safe experimentation.
• Open ecosystem: works with DuckDB, Spark, Ray, Hugging Face Hub, Unity Catalog and AWS Glue.

• Vietnamese AI startups managing image, video and text datasets for training vision and language models.
• Media and e-commerce platforms running multimodal search across product photos and videos.
• Autonomous vehicle and robotics teams curating sensor data for model training.

• One table for curation, features, search and training instead of separate pipelines.
• Lance format v2.2 cuts storage by over 50% and makes blob reads up to 68x faster, per the vendor.
• WeRide cut data mining from one week to one hour; Volcano Engine cut a pipeline from seven days to one.
• Up to 70% model FLOPS utilization when training directly from Lance tables.

AI teams increasingly train and serve models on video, images, audio and robotics data, which strains Parquet-based lakes and standalone vector stores. Demand for a single multimodal data layer is rising among autonomous driving, media and generative AI companies, including the growing AI and robotics developer community in Vietnam.

LanceDB Inc., based in San Francisco per its blog. Raised a 30 million USD Series A in June 2025. Released Lance SDK v1.0.0 in December 2025 and Lance format v2.2 in April 2026. Displays SOC, GDPR and HIPAA compliance badges.