Five Coding Agents Compared on One Local Qwen Model
Reddit user seti_at_home ran Claude Code, OpenCode, GitHub Copilot Desktop, Pi, and Codex once each with the same pixel-art Halloween prompt, using a locally served Qwen3.8 Flash Next model. The reported runs varied widely in elapsed time, request count, output tokens, prefill speed, and decode speed, while Pi encountered a memory stall and Codex was eventually aborted. A side-by-side video showed 30 seconds of each result, but the author did not score visual quality or conduct a controlled benchmark with identical token streams.
The self-reported run offers practical observations on how different coding-agent harnesses behave with the same local model, including their API formats, tool use, context handling, and operational overhead.
#Claude Code#OpenCode#GitHub Copilot Desktop#Pi#Codex#Qwen3.8 Flash Next#Halogen#Unsloth#AMD Ryzen AI Max+ 395