Gemini 3 Pro vs DeepSeek‑V3.2‑Exp: Which LLM Dominates SQL Understanding, Optimization, and Dialect Conversion?

Evaluation Overview

The SCALE benchmark evaluates professional‑grade large language models (LLMs) on database‑SQL tasks. This edition focuses on two models:

Gemini 3 Pro – released by Google on 2025‑11‑18, featuring strong reasoning, multimodal understanding, and code generation.

DeepSeek‑V3.2‑Exp – released on 2025‑09‑29, introduces a sparse‑attention mechanism for efficient long‑text processing and is open‑sourced on Hugging Face.

Methodology

The assessment follows the three core dimensions of the SCALE framework, using a unified test dataset to ensure fairness and reproducibility:

SQL Understanding – depth of logic, intent, and execution‑plan analysis.

SQL Optimization – ability to rewrite inefficient queries while preserving logical equivalence.

Dialect Conversion – accuracy of migration between different database dialects.

Gemini 3 Pro Detailed Results

SQL Understanding

Total Score: 86

Key Advantage: Highest execution accuracy, strong logical fidelity.

Improvement Area: Low execution‑plan detection and inconsistent structured‑output formatting (e.g., JSON null rendered as "NULL").

SQL Optimization

Total Score: 72.7

Key Advantage: Production‑grade safety (perfect syntax error detection) and high logical equivalence.

Improvement Area: Optimization depth could be deeper for complex patterns.

Dialect Conversion

Total Score: 77.1

Key Advantage: Strong performance on domestic database conversion.

Improvement Area: Limited capability for large‑SQL conversion and insufficient knowledge of specific domestic systems such as OceanBase.

Application Guidance

Target Users: Data analysts and database engineers.

Suggested Scenarios: Complex query logic verification, production‑grade performance tuning.

Value: Provides reliable data insights and safe code transformations.

DeepSeek‑V3.2‑Exp Detailed Results

SQL Understanding

Total Score: 66.7

Key Advantage: High syntax compliance.

Improvement Area: Weak execution‑plan detection indicates shallow understanding of database mechanisms.

SQL Optimization

Total Score: 61.5

Key Advantage: Strong syntax safety (≈90% error detection).

Improvement Area: Limited optimization depth for complex performance‑boosting strategies.

Dialect Conversion

Total Score: 58

Key Advantage: Outstanding domestic conversion capability.

Improvement Area: Poor large‑SQL conversion and low syntax‑error detection raise production risk.

Application Guidance

Target Users: Database engineers and enterprise decision‑makers.

Suggested Scenarios: Routine SQL syntax checks and initial domestic migration projects.

Value: Reduces manual effort for domestic migration but should not be used for deep performance tuning.

Common Technical Challenges

Execution‑plan parsing defects (e.g., mis‑labeling write‑operation plans, incorrect type fields).

Insufficient pattern recognition for optimization (e.g., failing to rewrite LIKE prefixes to range queries, missing redundant LIMIT / ORDER BY removal).

Type‑conversion awareness gaps (e.g., overlooking implicit DATE to string conversions).

Limited knowledge of specific domestic dialects, leading to syntax errors such as using unsupported functions or retaining vendor‑specific statements.

Resources

Benchmark repository:

https://github.com/actiontech/sql-llm-benchmark