Sebastian Raschka • 6/26/2026

Local Open-Weight LLMs in Coding Harnesses

This article evaluates local open-weight large language models (LLMs) such as Qwen-Code, Codex, and Claude Code in various coding harnesses. It highlights that 30B Mixture-of-Experts models offer a sweet spot, solving challenging problems at roughly 40 tok/sec on a Mac or DGX Spark, comparable to GPT 5.5. The author compares harness choices, noting Claude Code uses twice as many tokens as Codex, and includes Gemma 4 E2B as a reference to show smaller models struggle. A full write-up is available, with a chart comparing token use and task success across five local-agent tasks.

0 comments

#Local LLM #Model Comparison #Coding Agents