Agentic coding without the cloud: evaluating open-weight large language models on longitudinal data preparation tasks

Researchers benchmarked locally deployable open-weight LLM agents for longitudinal data preparation and found strong performance across 20 tasks and six cohort waves.

Top 31B–35B models reached near-saturated results, suggesting they can handle many routine steps with little performance loss.

The benchmark supports privacy-preserving, on-device AI assistance for research settings where cloud use is restricted.

Potential uses include multi-wave merging, category harmonization, and generating R code on consumer-grade hardware.