Local model coding explores a question that matters before anyone lets local models near real development work: are they actually good at coding, or are they just good at sounding confident in monospace? It asks local models to handle six common coding tasks: writing code, fixing bugs, completing an existing design, refactoring, generating tests, and reviewing code. Each answer is scored by a hosted GPT-5 evaluator against a detailed rubric, so the comparison is about coding behaviour rather than vibes.
The point is not that every benchmark prompt is heroic. It is that coding is a different skill from general language fluency. A model that writes a polished paragraph may still miss a boundary case, mutate input data, or produce tests that politely test nothing. This workbench separates those abilities and makes the failures visible.
The most useful result was practical: OpenAI’s smaller open-weight model performed best on a mid-level laptop with 16GB RAM, with scores approaching frontier online models. It was much slower, though Ollama performance was not tuned.