feyninc/sqrl-9b

Overview

feyninc/sqrl-9b is a 9-billion parameter agentic text-to-SQL model, derived from Qwen/Qwen3.5-9B. It specializes in translating natural language questions into SQL queries for SQLite databases. A key differentiator is its agentic protocol, which enables the model to perform read-only exploration queries to inspect database details like value formats or join paths before committing to a final SQL answer. This approach allows for more robust and accurate SQL generation, particularly for complex or ambiguous queries.

Key Capabilities

• Agentic Text-to-SQL: Generates SQL queries for SQLite databases, optionally performing intermediate exploration queries to gather information.
• High Accuracy: Achieves 69.8% execution accuracy on the BIRD-dev benchmark using a vote-of-8 sampling, demonstrating strong performance in its domain.
• Cost-Performance Sweet Spot: The 9B model's vote-of-8 performance is noted as the family's optimal balance between cost and accuracy, closely matching larger 35B models.
• Confidence Signal: Vote unanimity among samples provides a strong confidence signal, useful for routing or escalation of queries.
• Fine-tuned Training: Underwent SFT distillation from a stronger 35B teacher and online reinforcement learning (CISPO RL) with execution-match reward on cleaned BIRD and Spider datasets.

Good For

• Applications requiring accurate text-to-SQL generation for SQLite databases.
• Scenarios where database exploration (e.g., checking value formats, verifying joins) before final SQL generation is beneficial.
• Developers seeking a cost-effective yet high-performing text-to-SQL solution for agentic workflows.

Limitations

• Primarily tuned for SQLite and its specific identifier quoting conventions.
• Best used with the exact prompt protocol it was trained on.
• Generated SQL should be executed read-only and validated before acting on results.