A Thoughtworks technologist shares hands-on experiences running small language models locally on Apple Silicon (M3 Max and M5 Pro) for agentic coding tasks. The evaluation follows a viability funnel covering RAM fit, speed, tool calling, code correctness, context handling, and code quality. Models tested include Qwen3.6 35B MoE, Gemma 4 31B/26B, and Qwen Coder Next 80B MoE, using OpenCode and Pi as coding harnesses. Results were inconsistent — manual and automated evaluations often contradicted each other, and the same model produced dramatically different quality outputs on different machines. Day-to-day use with Qwen3.6 35B MoE showed it works best for small, well-defined tasks with minimal file discovery, especially in Bash and Python. The author concludes local models are far from matching large cloud models for agentic coding but offer a useful, more deliberate coding experience for constrained tasks.
