Lance
Lance is a free, self-hosted Apache-2.0 lakehouse format for multimodal AI data on object storage.
What Lance is
Lance is a free, self-hosted Apache-2.0 lakehouse format for multimodal AI data on object storage. It stores images, video, audio, text, embeddings, and tabular data together. Search combines vector similarity, BM25 full-text search, and SQL analytics on one dataset, using secondary indexes. The project reports 100× faster random access than Parquet and Iceberg. Its table format supports ACID transactions, schema evolution, time travel, dataset versioning, and incremental MVCC updates. Backfilled columns can be added through new Lance files instead of rewriting the whole table. Lance can commit directly to object storage, while Directory Catalog supports object-store deployment without extra infrastructure. Integrations include Pandas, Polars, Ray, PyTorch, DataFusion, DuckDB, Spark, Trino, Flink, Apache Polaris, Unity Catalog, Apache Gravitino, and Hive Metastore. The Python SDK is pylance, installed with `pip install pylance`. The unstable Lance 2.3 format has no compatibility guarantee and is not for production.
Good
- Stores multimodal and tabular data together
- ACID transactions and time travel
- Vector, text, and SQL search
- Direct object-storage commits
- Free Apache-2.0 license
Watch out
- Lance 2.3 is unstable
- No compatibility guarantee for Lance 2.3
- Documentation says Lance 2.3 is not for production
- Self-hosted deployment
Who it suits
Lance suits teams building multimodal AI datasets on object storage that need versioning, search, and SQL access. It is intended for self-hosted deployments and users comfortable with its listed integrations.
Verdict
Lance provides one format for multimodal data, transactional tables, and several search modes. Its broad ecosystem and object-storage focus are useful, but the unstable 2.3 format should be kept out of production.
Plans and prices
From Lance's own pricing page, checked 8 Oct 2026.
| Plan | Price | What it gives you |
|---|---|---|
| Lance | Free | Open source · Apache-2.0 |
The facts
Each one read from Lance's own pages, with the date it was checked.
What it does
- Catalogs
- Catalog implementations listed include Apache Hive, Apache Polaris, Apache Gravitino, Unity Catalog, and AWS Glue.checked 8 Oct 2026
- Community support
- The project links users to Discord and a mailing list for community communication.checked 8 Oct 2026
- Data management
- The table format supports ACID transactions, schema evolution, time travel, and incremental updates through MVCC.checked 8 Oct 2026
- Getting started
- The quickstart identifies the Python SDK, installed with `pip install pylance`, as the easiest way to start using Lance.checked 8 Oct 2026
- Incubating integrations
- The governance page identifies Lance Context, Data Viewer, Flink, Graph, Trino, and pglance as incubating projects whose APIs or functionality may change.checked 8 Oct 2026
- Integrations
- Documented integrations include DataFusion, Spark, DuckDB, Trino, PyTorch, TensorFlow, Ray, Hugging Face, PostgreSQL, and Flink.checked 8 Oct 2026
- License
- The site identifies Lance as open source under the Apache-2.0 license.checked 8 Oct 2026
- Multimodal data
- Lance stores images, videos, audio, text, and embeddings alongside tabular data in one format.checked 8 Oct 2026
- Notable limit
- The current unstable Lance 2.3 file format has no compatibility guarantee and the documentation says it should not be used in production.checked 8 Oct 2026
- Product
- Lance is an open source lakehouse format for multimodal AI, with file, table, and catalog specifications for object storage.checked 8 Oct 2026
- Purpose
- Lance is an open lakehouse format for multimodal AI, built for data on object storage.checked 8 Oct 2026
- Python SDK
- The quickstart identifies pylance as the Python SDK and gives `pip install pylance` as the installation command.checked 8 Oct 2026
- Random access
- The site says Lance offers 100× faster random access than Parquet and Iceberg.checked 8 Oct 2026
- Schema evolution
- Lance supports adding columns with backfill by writing new Lance files rather than rewriting the full table.checked 8 Oct 2026
- Search
- Lance combines vector similarity, BM25 full-text search, and SQL analytics on the same dataset, accelerated by secondary indexes.checked 8 Oct 2026
- Security reporting
- The community page asks users to report security vulnerabilities privately to the PMC mailing list and avoid public disclosure until the issue is resolved.checked 8 Oct 2026
- SQL access
- Lance datasets can be queried with SQL through integrations including DataFusion, Spark, DuckDB, and Trino.checked 8 Oct 2026
- Storage
- Lance can commit directly to object storage, and its Directory Catalog supports deployments directly on object stores without additional infrastructure.checked 8 Oct 2026
- Support
- The contributor guide directs users with questions to the Lance Discord for real-time support.checked 8 Oct 2026
- Versioning
- The format supports ACID transactions, time travel, and dataset versioning.checked 8 Oct 2026
Alternatives to Lance
- 1
Roboflow
Free planAndroidiOSWebAPI8.3From $39 a month - 2
Weights & Biases
Free planAndroidiOSWebAPI8.3From $60 a month - 3
DagsHub
Free planWebAPILinuxmacOS7.9From $99 a month - 4
git-annex
Free planOpen sourceAndroidWebAPILinux7.7Free plan - 5
ClearML
Free planWebAPILinuxmacOS7.4From $15 a month - 6
Neon
Free planWebAPILinuxmacOS7.1$0.11 usage-based