Experiences with local models for coding

This article by Birgitta Böckeler, a Distinguished Engineer at Thoughtworks, details her hands-on experiments with running small language models locally for agentic coding tasks. She describes a 'viability funnel' process: checking RAM fit, response speed, tool calling ability, code correctness, and handling of complex tasks. Using M3 Max and M5 Pro machines, she shares her journey through manual and automated evaluations, highlighting the challenges of model configuration, internet speed, and the iterative testing needed to assess code quality and user experience. The article is part of a series on generative AI for software development.