Can a local AI write code?
Yes, within limits. The honest size of those limits, and the reason the work stays on your disk.
Yes, within limits. Small open models write working code from a plain description, fix bugs, and explain unfamiliar files, because coding is a core language task they were trained on. The quality depends on model size and the difficulty of the task: small models do well on everyday scripts and struggle at large or novel problems. Reviewing the output stays essential, and nothing you paste ever leaves your machine.
Coding is a language task.
There is nothing magical about a model producing code. Code is text with strict conventions, it appears in enormous quantities in the text models train on, and predicting it well is the same skill the model uses everywhere else. That is why even compact local models can emit a working script from a plain description, translate between languages, or spot the bug in a snippet you paste. It also explains the limits: the model reproduces patterns it has seen, so the closer your task is to common, well-documented code, the better it does.
What small models handle, and what they do not.
The everyday work goes fine. Scripts under a page, data-format conversions, regular expressions, configuration files, explanations of an unfamiliar function: these are bounded tasks with clear inputs, and models in the 7B to 8B range manage them routinely; what that size means is a memory question before it is a quality one. The work that strains them is long-horizon reasoning: a feature spanning many files, an algorithm nobody has written down, a refactor that has to keep a whole system consistent in its head. A local model can attempt these and should not be trusted with them unreviewed, which is true of bigger models too but bites harder here.
Where the local version wins.
For a lot of code work, the privacy is the feature. Pasting internal source into a cloud assistant means sending proprietary material to someone else's machine, a trade many employers now forbid. A local model makes that trade unnecessary: the model runs on your hardware, the code never travels, and nothing is logged anywhere you cannot see. The workflow is honest about the trade anyway, since generated code is a draft with confident errors possible in it, so the review step is not optional either way. Local buys you the ability to iterate freely on code you could not paste anywhere, at the cost of capability on the hardest problems.