Comparison of LLM Models

from /villagepump/2024/04/30

https://artificialanalysis.ai/

Wait, isn’t the 3rd party provider for LLaMA 3 70B too cheap?

• It seems to deliver performance close to GPT4 and Claude 3 Opus but at a price that is less than 1/10th.
  • Performance-wise, it feels similar to Claude 3 Sonnet.
• 
• I wasn’t aware that things had turned out like this (blu3mo).
• So, if you want to use the best Open Source model right now, what should you go for?
  • I hear names like Mistral, Llama, Phi being mentioned.
  • If you want to generate Japanese output, should you use a fine-tuned model?
    • In that case, which infrastructure should you use?

  • When enhanced versions of phi3 or llama3 for Japanese learning come out, the world is likely to change significantly. Seriously, people developing products should at least prepare to be able to adapt to ollama’s npm or REST API, or things could get pretty messy.

  • @(erukiti) [April 23, 2024](https://twitter.com/erukiti/status/1782908698513612897?ref_src=twsrc%5Etfw)
    

  • Command R+, strong, Japanese-friendly, and relatively inexpensive?
• This is exactly the kind of thing I want AI to research… (nishio)
  • I don’t feel motivated to research differences between models or differences between machines and GPUs.