DuckDB 2.0.0 LTS, codenamed Cinnamon Teal, is released as the biggest feature release to date, introducing a polished Quack client-server protocol, asynchronous I/O, and a first-class VARIANT type, with long-term support through at least October 2027.

The 2.0.0 launch ships with nearly 20,000 commits since 1.5 and brings major features including a new PEG-based SQL parser, an agent mode for the CLI, and enhanced extensibility, along with dedicated support through late 2027.

Users are warned that upgrading to 2.0 creates files incompatible with pre-2.0 versions, so migration should be tested on a shadow system, with the first patch due in November for mission-critical workloads.

DuckDB now offers three release lines: 2.0 as the long-term support version through at least October 2027, 1.5 as an interim LTS through March 2027, and 1.4 as an older LTS supported until late 2026.

Storing JSON as VARIANT yields storage savings of about 27% in tested setups, though performance gains depend on workload, with some naive cast-and-query scenarios not showing speedups.

A revamped C API v2 enables hosting and signing extensions and introduces a framework for dynamic extension repositories that DuckDB can load.

Triggers are supported with BEFORE/AFTER and FOR EACH ROW/STATEMENT, including transition tables via REFERENCING OLD/NEW TABLE, though some cases like BEFORE INSERT FOR EACH ROW are not yet implemented.

The new CONNECT statement lets users attach and switch to remote databases, enabling remote pushdown and server-client workflows via Quack.

SQL dialects are expanded with variables using $ shorthand, data-modifying CTEs, nearest joins for vector workloads, nested schemas, and in-place JSON mutation functions.

A new PEG-based SQL parser replaces the old PostgreSQL-derived parser, improving extensibility and error precision while remaining compatible to ease migration.

Asynchronous I/O improves performance for reading Parquet and CSV files by enabling background reads and parallelism, boosting performance on object storage such as S3.

CLI now includes agent mode for optimized output to AI agents, with Markdown tables, cost estimates, and JSON error reporting, and it can be toggled on or off.