Feyn AI Releases SQRL, a Text-to-SQL Model Family That Inspects the Database Before Writing a Query
What changed
Feyn Labs launched SQRL, a family of text-to-SQL models that query databases more cautiously. Before generating SQL queries, SQRL inspects the database schema and samples data through read-only probes. The flagship model, SQRL-35B-A3B, achieves 70.6% execution accuracy on the BIRD Dev benchmark, slightly outperforming Claude Opus 4.6. Feyn also offers distilled, self-hostable versions in 4B and 9B parameter sizes, enabling deployment without cloud dependency.
Why builders should care
Text-to-SQL tasks typically rely on models generating queries solely from user prompts. SQRL’s approach reduces errors by understanding database structures upfront instead of guessing based on prompt context alone. This pre-inspection can prevent malformed queries, faulty assumptions, and runtime errors that disrupt workflows or cause data access issues. Builders using SQL automation gain better accuracy and robustness in complex, real-world environments where schemas vary and evolve.
The practical takeaway
Including a read-only database inspection stage before query generation makes text-to-SQL models more reliable for applications like analytics dashboards, automated reporting, and natural language interfaces to databases. The self-hostable 4B and 9B versions let companies run these models on-premises, which helps with data privacy, latency, and cost control. Operators should evaluate SQRL as a more precise and flexible alternative to other text-to-SQL tools that do not verify schema or content beforehand.
What to watch next
Pay attention to how SQRL’s probing approach influences developer expectations for text-to-SQL APIs, especially in enterprise settings. Its ability to self-host smaller checkpoints might accelerate adoption in regulated industries wary of cloud usage. Future updates may refine the balance between inspection depth and query latency. Also watch if competitors adopt similar pre-query inspection techniques or raise the bar for execution accuracy on standardized benchmarks.
AI Quick Briefs Editorial Desk
