241
The circle of life (thelemmy.club)
submitted 3 months ago by to c/programmerhumor@lemmy.ml
13 comments fedilink hide all child comments
 
you are viewing a single comment's thread
view the rest of the comments
[–] 13 points 3 months ago (5 children)

One jas to Knox how expensive LLMs are by trying to run one at home. Tried running it myself, figured out I need more than 128GiB of RAM and thousand dollars in grahics card. Figured out a $5 openai per month is cheaper, and also understood they are burning money by providing a free service and 20$ subscription.

  • permalink
  • fedilink
  • source
  • hideshow 5 child comments
    •
  • [–] [S] 11 points 3 months ago (4 children)

    You should be able to get very decent performance with 128gb vram running Qwen 3.6 with something like https://github.com/itigges22/ATLAS especially if you run MTP https://huggingface.co/unsloth/Qwen3.6-27B-MTP-GGUF

    A friend of mine gets something like 50 tokens a second with it, and output quality is quite decent.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 4 child comments
    •
  • [–] 10 points 3 months ago (1 child)

    Yeah, but depending on your location (and usage), you might be burning more than $5/mo in electricity to run that shit. Not to mention the costs of buying all that hardware ... especially at current inflated rates.

    If you have to buy 128GB of RAM in 2026, it's going to be a long time before you come out ahead vs paying $20/mo for some AI subscription.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] 3 points 3 months ago (1 child)

    How does it compare to largest deepseek ans Claude opus 4.6? I hot used to blazing fast speed and accurate results. I'm not buying a server and 128 GB of RAM just to run a model similar to gpt-4.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] [S] 5 points 3 months ago

    ATLAS has some benchmarks in the repo, and it's comparable to opus 4.6, you don't actually even need 128gb model for that. An 8 bit quantized model will run with around 32gb and still perform quite well.

  • permalink
  • fedilink
  • source
  • parent