Feyn AI Releases SQRL Text-to-SQL Models That Inspect Databases

Feyn AI released SQRL, a new family of text-to-SQL models, on an unspecified date, reframing the task from simple translation to one involving database inspection. Unlike traditional text-to-SQL systems that generate queries immediately, SQRL first inspects the database to resolve ambiguities and ensure the generated queries are supported by the actual data. This approach aims to prevent common errors such as incorrect joins, misinterpretation of ambiguous columns, or filtering for non-existent values, which often go undetected by execution alone.

The flagship model, SQRL-35B-A3B, has demonstrated a 70.6% execution accuracy on the BIRD Development (Dev) dataset. This performance surpasses that of Claude Opus 4.6, which achieved 68.77% accuracy under the same evaluation conditions. The BIRD benchmark is designed to measure text-to-SQL performance on real-world domains, incorporating imperfect values, ambiguous columns, and complex relationships, thereby assessing not just syntactic correctness but also the accuracy of returned results.

Feyn AI's core innovation lies in enabling the model to query the database for clarification before formulating a response. SQRL receives a natural language question, the database schema, and optional evidence about the database. If the provided context is sufficient, it generates a query directly. However, if ambiguity persists, the model executes read-only queries to gather information from the database, which it then uses to draft the final SQL query. This inspection phase is crucial for ensuring the query's relevance and accuracy.

Three checkpoints of the SQRL model family are available on Hugging Face: SQRL-4B, SQRL-9B, and SQRL-35B-A3B. These releases provide researchers and developers with access to different scales of the model, allowing for experimentation and integration into various applications. The company, a YC-backed startup, emphasizes that schema information alone is insufficient to prevent errors, as it does not clarify data formatting or potential join complexities.