Studio10 coding agents tested on local Qwen 3.8 27B: Pi uses only 9.2k tokens per task and passes 5/5
Original title: Which coding agent gets the most out of a local Qwen 3.8 27B? I lined up 10 and logged tokens, cache and sandbox blocks
The author ran Qwen 3.8 27B Q4 (thinking off) via llama.cpp on a MacBook Pro M5 Max 64GB and benchmarked 10 coding agents on 5 trap-laden coding tasks in a sandbox, logging tokens, cache reuse and internet attempts. Pi was leanest at 38s median and about 9.2k tokens per task, while Qwen Code was hungriest at about 107.9k. Atomic Agent was the only one using llama.cpp's native endpoint with grammar-constrained tool calls, reusing 86% of cache and passing 5/5 fully offline. Codex, Cline, goose and OpenCode tried to reach the internet every run and were blocked by the sandbox, and OpenCode, goose and Atomic Agent miscounted 244 lines as 64 ERRORs in the log task. The author discloses working on Atomic Agent.
The author tested 10 coding agents on the same local model and sandbox, producing comparable data on tokens, cache reuse and internet attempts that readers can use to compare their own tools.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.