SFT + simple search tool lets Qwen3-4B match a ~1T-param model on diagnosis benchmark

Tim_Dettmers · x · 2026-10-18

Adarshsolanki spent days training a tiny model to diagnose patients in CMU's Synthetic Hospital environment: RL helped little, but some SFT plus a simple search tool got Qwen3-4B to perform on par with a 1T-param model (GPT-6 Luna). Tim Dettmers highlighted the key insight: tools help small models but not large ones, meaning capability gaps can be replaced by tools — the agent is more than just the model.

Original post →

SFT + simple search tool lets Qwen3-4B match a ~1T-param model on diagnosis benchmark

More from coding & agent