4B Model Beats Postgres by 81%: What a $1,200 Training Run Means for Philippine AI

A researcher trained a 4-billion-parameter Qwen model on two used RTX 3090s for $1,200: it builds database query plans 81% faster than Postgres. The LoRA adapter with 21.2 million parameters weighs 42.5 MB and fits on a smartphone.

• Training: ~$800 for 2x H100 rental for 95 hours and ~$400 on API trajectories
• Result: 1.81x speedup and 44.7% reduction in total latency
• LoRA adapter: 21.2 million parameters, 42.5 MB, fits on phone
• RL method based on GRPO from DeepSeekMath 2024 paper