Best LLMs for SQL Generation: Local Database Assistants Ranked

SQL generation is one of the highest-value use cases for local LLMs. Running it locally means your database schemas and sensitive data never leave your machine. This guide benchmarks the top locally-runnable models on SQL generation tasks.

SQL Benchmark: Spider and Bird

Model	Spider (EX)	BIRD (EX)	VRAM
Qwen3.6 32B	89.3%	72.1%	20 GB
DeepSeek-Coder 33B	88.1%	70.4%	20 GB
Llama 3.3 70B	87.8%	69.9%	42 GB
GPT-4o (cloud)	89.6%	73.8%	API

Setup: SQL Assistant with Schema Context

import ollama

SCHEMA = """
CREATE TABLE gpu_benchmarks (
  id INTEGER PRIMARY KEY,
  model_name TEXT,
  vram_gb INTEGER,
  tokens_per_sec FLOAT,
  price_usd INTEGER
);
"""

def query_to_sql(natural_language_query: str) -> str:
    response = ollama.chat(
        model='qwen3.6:32b',
        messages=[
            {'role': 'system', 'content': f'''You are a SQL expert. Generate PostgreSQL queries only.
Database schema:
{SCHEMA}
Return only the SQL query, no explanation.'''},
            {'role': 'user', 'content': natural_language_query}
        ],
        options={'temperature': 0}
    )
    return response['message']['content']

# Example
sql = query_to_sql("Which GPUs have over 20GB VRAM and cost under $1000?")
print(sql)
# SELECT model_name, vram_gb, price_usd FROM gpu_benchmarks 
# WHERE vram_gb > 20 AND price_usd < 1000;

For the hardware setup needed to run 32B models for SQL generation tasks, see LLMHardware.io's developer build guide.