Qwen3.5-0.8B reportedly improves from 10% to 36% on a test set

A user says they improved Qwen3.5-0.8B from 10% to 36% on their test set by reducing tool calls by almost a factor of two. They say adapting a small model to a target harness with reinforcement learning and Hugging Face jobs can cost under $20, fit on one 24GB GPU and require no synthetic trajectories.