• favoredponcho@lemmy.zip
      link
      fedilink
      English
      arrow-up
      5
      ·
      4 hours ago

      To use one equivalent to Claude Opus, you need like 800GB GPU memory. The chips that will get you there run $20-30k a pop, and you’d need 4 or more of them.

      • boonhet@sopuli.xyz
        link
        fedilink
        English
        arrow-up
        2
        ·
        2 hours ago

        That’s the biggest baddest model out there. There are models that get you 90% of the way there with a significantly smaller parameter count and thanks to MoE offloading you don’t need the entire model active at once.

        A 5090 and a beefy CPU and tons of RAM won’t be cheap or even affordable to most, but you could run very big models and have a beefy PC for other activities. But even a 16 GB card could do plenty.

      • kamen@lemmy.world
        link
        fedilink
        English
        arrow-up
        3
        ·
        4 hours ago

        Yeah, I’m aware. I have realistic expectations and I’m looking into running something simpler and less demanding.

    • melfie@lemmy.zip
      link
      fedilink
      English
      arrow-up
      5
      ·
      edit-2
      7 hours ago

      Where I work, Chinese models are banned due to legal concerns. Not just in production, on any company-owned machine. That basically eliminates all the decent open weight models. I’m imagining this type of policy will be more widespread. I suppose it’s because of the potential for legal woes if systems and people are dependent on these models and then federal or state laws impose harsh penalties with little time to react.