It technically handles 1M tokens, but if you ask it questions it's obvious that it's too much to handle.

Just upload a novel and ask it questions, you'll see how it botches simple stuff

Easiest way to prove it can't handle the full context in reality is to upload a one hour documentary movie with audio and ask it to write timestamps of chapters/critical moments. It can't handle this beyond 10 minutes even remotely reliably.

Good or bad it's a zillion times better than Claude or ChatGPT where you can't even upload a zipfile.

If you want this in numbers check the nolima benchmark