• wise_pancake
    link
    fedilink
    English
    arrow-up
    7
    arrow-down
    4
    ·
    2 days ago

    Local AI kind of sucks right now, and everything is massively over branded as AI ready these days.

    There aren’t a lot of compelling local use cases and the memory constraints of local mean you end up with fairly weak models.

    You need a high end high memory local setup to get decent token rates, and I’m finding right now 30-70b models are the minimum viable size.

    That doesn’t compare with speed of online models running on GPUs that cost more than luxury cars.