Linux

That's the neat part, you can.

If I had to pick one, it'd be Olmo 3 from Ai2: Non-profit. Ai2 is a research institute, not a company that needs to sell you a subscription or harvest your prompts. Fully open, not just "open weights" Ai2 publishes the training data, the training code, every checkpoint and the logs. Anyone can check exactly what went in, and in principle rebuild it from scratch. You don't have to take their word for anything.

Runs locally. The 7B version runs on an ordinary gaming laptop or a recent MacBook. Your code never leaves your machine, there's no data centre involved when you use it, and the energy cost is your laptop being on. Apache 2.0 licence. No usage restrictions and no strings attached.

Good enough for the tutor use case. Explaining errors, walking through why code misbehaves, and pointing out likely bugs are all well within what it can do.

To be fair about it: the training data is largely scraped from the web, like nearly every model. The difference is that Olmo is honest and auditable about it.

If strict consent matters most to you, StarCoder2 is cleaner (only permissively licensed code, with an opt-out)