Small open-weights model reports faster PostgreSQL plans in benchmark test

A 4-billion-parameter open-weights model was trained to generate hints for PostgreSQL query plans, with the experiment’s author reporting a 1.81× geometric mean speedup and a 44.7% reduction in summed latency across the Join Order Benchmark.

Rohan Bansal detailed the experiment in a report linked through Hacker News. The model was trained with supervised fine-tuning on trajectories generated by larger models, then refined through agentic reinforcement learning based on measured PostgreSQL execution times.

The system used the third-party pg_hint_plan extension to test alternative join orders, join algorithms and scan methods. Training and testing used the IMDb database, with the Cardinality Estimation Benchmark used for training and the 113-query Join Order Benchmark reserved for evaluation. Query topologies were separated between the datasets.

The reported final result came from three trajectories per query, with the best candidate selected from up to 15 generated plans. That best-of-15 process is different from single-shot model performance and reflects a proposed use case in repeated analytical workloads, where the cost of upfront experimentation could be spread across many executions.

The experiment also found that increasing PostgreSQL’s shared buffers from 128MB to 2GB reduced measurement noise and cut the benchmark workload’s summed runtime from 95 seconds to 60 seconds. Bansal reported spending about US$800 on H100 compute and US$400 on API fees.

The results are limited to the author’s benchmark setup and were not independently verified in the supplied material. They show PostgreSQL being steered towards alternative plans, rather than its optimiser being replaced, and do not establish performance gains across other databases, schemas, workloads or configurations.