Comment by jay_kyburz
4 days ago
OK, for those interested, I have a Ryzen9 2700 with a 12GB RX6750XT and downloaded Gemma3:4b because that was what the app suggested first.
The speed seems fine to me, but the hallucinations are wild, completely wrong on a few things I like to test the commercial offerings on.
For simple questions about the lua language and how to do things in Unity game engine the results look fairly OK.
Honestly to be expected with a 4b model. 12b/14b+ is the minimum in my experience to get decent results, unless you have a specific use-case for the 4b ones and fine-tune it to your use.
Yeah, I tried Deepseek 8b as well and it was hopeless, but it was interesting to watch it think. I haven't seen that before.