Feyn AI has released SQRL, a family of text-to-SQL models that inspect database contents via read-only probes to resolve ambiguity before generating final queries. The flagship SQRL-35B-A3B achieves 70.6% execution accuracy on the BIRD Dev benchmark, outperforming Claude Opus 4.6.

  • Three open checkpoints are available: SQRL-4B, SQRL-9B, and SQRL-35B-A3B, all built on Qwen3.5/3.6 families.
  • The model uses a mixed strategy of single-shot generation for clear questions and up to five inspection rounds for ambiguous ones.
  • Training involved cleaning BIRD and Spider datasets, using CISPO reinforcement learning, and distilling teacher trajectories into student models.
  • SQRL-9B is recommended as the default checkpoint, serving via vLLM with a read-only harness that feeds observations back to the model.

The approach allows users to host self-hostable models that maintain high accuracy while keeping schema and query data on private infrastructure.