AlioLeuchtmann/qwen-3.5-sql-27B-Distill-2B
AlioLeuchtmann/qwen-3.5-sql-27B-Distill-2B is a 2.3 billion parameter text-to-SQL model distilled from a Qwen 27B teacher, specifically fine-tuned for SQLite text-to-SQL tasks. It achieves 68.47% execution accuracy on the official BIRD benchmark, outperforming models significantly larger in size. This model excels at generating SQL queries from natural language questions, particularly with execution-verified structured chain-of-thought distillation.
Loading preview...
Model Overview
AlioLeuchtmann/qwen-3.5-sql-27B-Distill-2B is a compact yet powerful 2.3 billion parameter model specialized in text-to-SQL conversion for SQLite databases. It stands out as the smallest model to achieve top-tier performance on the BIRD single-model leaderboard, demonstrating that knowledge distillation can effectively compress complex task competence into smaller models.
Key Capabilities
- Exceptional Text-to-SQL Performance: Achieves 68.47% Execution Accuracy (EX) on the official BIRD test set, rivaling and often surpassing much larger general-purpose LLMs and even some specialized models.
- Efficient Distillation: Trained using execution-verified structured chain-of-thought distillation from a Qwen 27B teacher, without reinforcement learning or extensive test-time reasoning.
- Resource-Efficient: Its 2.3B parameter size allows it to run on consumer GPUs or edge hardware, offering significant cost and latency advantages for specialized tasks.
- Structured Output: Generates a structured analysis followed by a SQL query, designed for non-thinking decoding.
When to Use This Model
- SQLite Text-to-SQL Applications: Ideal for applications requiring accurate SQL query generation from natural language questions against SQLite databases, especially with BIRD-style analytical questions and evidence hints.
- Resource-Constrained Environments: Suitable for deployments where computational resources are limited, such as on-premises or edge devices.
- Verifiable Domains: Demonstrates the potential of distillation for domains where task outputs can be execution-verified, suggesting applicability to similar verifiable tasks.
Limitations
While highly effective for its specialized domain, this model is not instruction-tuned for general chat. Its performance on highly challenging, multi-step reasoning tasks still trails larger models.