I used to think local AI was mostly useful for demos and small experiments.

Then I spent a few days testing Gemma 4 on my laptop.

The interesting part wasn’t benchmark numbers. It was how usable the model actually felt during normal development work.

I tested it with:

Rails services
old migrations
background jobs
messy business logic

and it handled repository context better than I expected from a local model.

It still struggles with bigger architectural decisions and long autonomous tasks, but debugging and code understanding were surprisingly solid.

The biggest takeaway for me:
local models are finally becoming practical enough that I’d actually keep one running during daily work.