• Ptsf@lemmy.world
    10·
    1 day ago

    You literally can. Kimi k3 is fist for fist punching at both sol and fable levels. The smaller qwen 3.8 27b is benching well against gpt 5.4 and opus 4.7. There are unreleased internal models you cannot run locally but you also can’t buy them. Additionally, big models are generally bad tools, you’re much better off tuning a local model for your specific task (Wendel from level 1 has a good video on this). If you’re gonna make claims at least do a Google search.

    • flamingos-cant (hopepunk arc)@feddit.ukEnglish
      4·
      1 day ago

      If you want something useable you need a decent amount of RAM/VRAM and a modern GPU, things that aren’t exactly accessible these days. That Kimi model is apparently half a terabyte in size and requires 100s of GB of VRAM, so I don’t know why you’re saying it’s a ‘local’ model.

      • Ptsf@lemmy.world
        10·
        1 day ago

        That is patently untrue. You can have useful ai on a esp32, a $10 chip. https://github.com/slvDev/esp32-ai It will not be as capable as kimi k3 in general generic tasks but you can fine tune models for certain workflows (literally how “distillation” works) and you can run them on affordable and available hardware. Just because you don’t understand a tool or how to properly use it does not mean someone else doesn’t.

        • flamingos-cant (hopepunk arc)@feddit.ukEnglish
          5·
          24 hours ago

          Just because you don’t understand a tool or how to properly use it does not mean someone else doesn’t.

          You brought up the Kimi model up in a conversation about local models being viable, something that requires 8 Nvidia GB300s as a prerequisite. And I’m sorry I don’t follow you’re little hobby close enough for you, I did try it and it sucked.

          • Ptsf@lemmy.world
            5·
            24 hours ago

            Kimi k3 can be ran at a cost of $37,900 with Mac studios. It’s not a hobby. Cool, glad for you. Tools are tools, I bet you don’t own or like an absolute gob of them but it doesn’t change their usefulness, and you don’t need kimi k3 for everything, I mentioned it specifically because it’s comparable to generalized frontier ai models and can be ran locally, but I also stated that you’re better off tuning models for your workloads specifically. Jfc why even comment if you’re gonna be so dumb and selective with your reading?

            • flamingos-cant (hopepunk arc)@feddit.ukEnglish
              3·
              23 hours ago

              I bet you don’t own or like an absolute gob of them but it doesn’t change their usefulness

              I was commenting based on my experience with this stuff, it was a lot of work to set up and the results I got were bad. I never even contended that it could be useful (but I’m reading selectively), I said you that you need decent hardware for it to be useable. I have an RX 6800 XT and 32GB of RAM and it was really slow, even with recommended models.

              Jfc why even comment if you’re gonna be so dumb and selective with your reading?

              I tried to engage in good faith, and you proceeded to call me an idiot, sorry I reacted poorly to you being rude. You get what you put in.

              • Ptsf@lemmy.world
                3·
                23 hours ago

                Nah brother I’m not even going to continue giving you my time. You called my science and field a hobby, you questioned my reasoning when it was already presented, and you did not approach in good faith, you approached with preconceived notions, prejudice, and shit opinions based on your bad configs. I don’t owe you shit.