On-device text-to-SQL distilled from GPT-4o-mini into Qwen2.5 (0.5B → 3B locally on M1 via mlx-lm LoRA, 7B+ on cloud A100). 847 MB at 62.5% on Spider dev; 3B variant hits 72.6%, 7B variant hits 75.0%.

Three-stage Text-to-SQL framework achieving 70% execution accuracy on Spider benchmark — runs free on any laptop with no GPU required