Sebastian Raschka • 6/26/2026

Local Open-Weight LLMs in Coding Harnesses

This article evaluates local open-weight LLMs (e.g., Qwen-Code, Codex, Claude Code) in coding harnesses, focusing on 30B Mixture-of-Experts models that achieve ~40 tok/sec on Mac or DGX Spark, comparable to GPT 5.5. It compares token efficiency and task success across five local-agent tasks, noting Claude Code uses twice as many tokens as Codex. Includes a chart and reference to a full write-up.

0 comments

#Local LLM #Coding Agents #Open Weight Models