feyninc/sqrl-35b-a3b

Overview

feyninc/sqrl-35b-a3b is a 35.1 billion parameter Mixture-of-Experts (MoE) model, with around 3 billion active parameters, designed for agentic text-to-SQL tasks. It is the flagship model of the sqrl family and served as the teacher model for smaller distilled versions like sqrl-4b and sqrl-9b. The model's core capability is to answer natural-language questions by generating SQL queries for SQLite databases.

Key Capabilities

• Agentic Text-to-SQL: Translates natural language into SQL for SQLite, with an optional exploration phase.
• Database Probing: Can run read-only exploration queries to inspect database schema, value formats, and join paths before committing to a final SQL answer.
• RL-Sharpened Performance: Achieves strong performance on the BIRD-dev benchmark, with 68.7% pass@1 (average over 8 samples) and 70.6% vote-of-8 (execution-clustered).
• High Unanimity: Answers 1,072 out of 1,534 questions unanimously, indicating high single-shot reliability, though with less sample diversity.
• Confidence Signal: Unanimous answers are right ~87% of the time, providing a strong confidence signal for routing or escalation.

Training and Protocol

The model was trained using CISPO RL (group-relative advantage) directly on the base model, utilizing async dynamic sampling on a difficulty-filtered subset of cleaned BIRD-train questions. It employs a specific agentic protocol where it emits exactly one action block per turn, either an <sql> exploration query or an <answer> final SQL query, after a brief reasoning summary. The model is heavily optimized for this protocol, and its performance is best when used with the specified prompt format and agentic loop.

Intended Use

This model is specifically built for SQLite text-to-SQL tasks, leveraging schema and optional evidence provided in context. It is not tuned for other SQL dialects and follows SQLite backtick conventions for identifier quoting. Users should always validate and execute generated SQL in a read-only environment first.