Big day for people who use AI locally. According to benchmarks this is a big step forward to free, small LLMs.

      • brucethemoose@lemmy.world
        link
        fedilink
        English
        arrow-up
        0
        ·
        edit-2
        4 months ago

        At long context (close to the full 128K), Nemo is way better than llama 8B in my testing.

        Turns out they are both very sensitive to quantization though.

        TBH I didn’t know people here were running LLMs. Seems like most of Lemmy is very broadly anti AI?

        • ObsidianZed@lemmy.world
          link
          fedilink
          English
          arrow-up
          0
          ·
          4 months ago

          My impression is the general consensus is we don’t want huge corporations stealing data to train their AI models only to turn around and cram it down our throats anywhere they can with increasingly negative experiences. That being said, while I would generally agree with that, I still find it interesting and especially if I can host it myself.

        • Bilb!@lem.monster
          link
          fedilink
          English
          arrow-up
          0
          ·
          4 months ago

          If forced to characterize the attitude of lemmy towards LLM/“AI,” I’d say people here are broadly interested in the tech but critical of the way it’s often used.

          • General_Effort@lemmy.world
            link
            fedilink
            English
            arrow-up
            0
            ·
            4 months ago

            If by interested you mean willing to bullshit… Talking about AI here is like talking about evolution at bible camp in the deep south.

          • brucethemoose@lemmy.world
            link
            fedilink
            English
            arrow-up
            0
            ·
            edit-2
            4 months ago

            I dunno, with image models specifically it seems like they’re the devil because of the datasets they’re trained on, killing artists, and… that’s that. And LLMs to a lesser extent. There’s truth to all that, but there’s also a lot more.

            I think most people don’t realize how much of an inflection point local running vs. corporate hosting could be, which is especially ironic on Lemmy.

        • just another dev@lemmy.my-box.dev
          link
          fedilink
          English
          arrow-up
          0
          ·
          4 months ago

          Yeah, there’s a massive negative circlejerk going on, but mostly with parroted arguments. Being able to locally run a model with this kind of context is huge. Can’t wait for the finetunes that will result from this (*cough* NeverSleep’s *-maid models come to mind).

          • brucethemoose@lemmy.world
            link
            fedilink
            English
            arrow-up
            0
            ·
            edit-2
            4 months ago

            I am looking into doing it on the 12B for myself, not so much for RP but novel style prose.

            I am thinking literature + a fanfic dump as a dataset?

            • just another dev@lemmy.my-box.dev
              link
              fedilink
              English
              arrow-up
              0
              ·
              4 months ago

              Ah, that’s a wonderful use case. One of my favourite models has a storytelling lora applied to it, maybe that would be useful to you too?

              At any rate, if you’d end up publishing your model, I’d love to hear about it.

                • just another dev@lemmy.my-box.dev
                  link
                  fedilink
                  English
                  arrow-up
                  0
                  ·
                  4 months ago

                  Oof - not on my 12gb 3060 it doesn’t :/ Even at 48k context and the Q4_K quantization, it’s ollama its doing a lot of offloading to the cpu. What kind of hardware are you running it on?

                  • brucethemoose@lemmy.world
                    link
                    fedilink
                    English
                    arrow-up
                    0
                    ·
                    edit-2
                    4 months ago

                    A 3090.

                    But it should be fine on a 3060, with zero offloading.

                    Dump ollama for long context. Grab a 5-6bpw exl2 quantization and load it with Q4 or Q6 cache depending on how much context you want. I personally use EXUI, but text-gen-webui and tabbyapi (with some other frontend) will also load them.

      • just another dev@lemmy.my-box.dev
        link
        fedilink
        English
        arrow-up
        0
        ·
        4 months ago

        I haven’t given it a very thorough testing, and I’m by no means an expert, but from the few prompts I’ve ran so far, I’d have to hand it to Nemo concerning quality.

        Using openrouter.ai, I’ve also given llama3.1 405B a shot, and that seems to be at least on par with (if not better than) Claude 3.5 Sonnet, whilst being a bit cheaper as well.

        • brucethemoose@lemmy.world
          link
          fedilink
          English
          arrow-up
          0
          ·
          4 months ago

          Llama 70B is probably where its at, if you go the API route. It’s distilled from 405B, and its benchmarks are pretty close.