Phi3 first thoughts

Posted by fab_space@reddit | LocalLLaMA | View on Reddit | 2 comments

As many of us I am testing wherever I can, in my case M2 metal laptop and a dell r630. Running it over the Dell via Ollama it is insanely faster and accurate than any other small model i tested. Can’t wait to test on laptop and try to write some agents 🤪

Reply to Post

2 Comments

[-]

cyan2k@reddit

It's impressive and amazing for structured output. currently letting phi3 creating backlogs for projects. literally hundreds of json objects. not a single error in generating json. not a single hallucinated or duplicated task or random non-sense tasks. and all without RAG or any kind of magic except some agents prompting each other. Most 8B models fail sometimes. Phi3 seems very stable. Even ChatGPT 4 was impressed and said it couldn't have done it better, haha https://imgur.com/a/C9pGFOa

[-]

Rufusthe13thapostol@reddit

What's the hardware requirements?