A 4B open-weights model was post-trained via supervised fine-tuning and a custom GRPO reinforcement learning variant to generate Postgres query plans.

It achieved a 44.7% latency reduction across 113 join-heavy queries after initially failing to produce valid plans for 99 of them. The RL loop scores rollouts by execution time, measured on a Postgres rig designed to minimize Linux page cache contention across concurrent containers.