Reddit user benchmarks 4 agent frameworks with identical model and prompt

View on Reddit

A single self-hosted Qwen3.6-27B (Q4) model on llama.cpp was tested with 4 different agent scaffolds (pi, opencode, hermes, qwen code) using the same prompt and hardware. The task was a single-file 2D canvas solar system with scripted orbits and gravity.