Insane performance
I'm using the F16 version on LM Studio.
This is insane, haha. I think.. CN AI has super-capable AIs ready.
They always make their SOTAs perform a bit less than the very best US AI models (on purpose) because the fact this 3B can do what it does.. shows they know what to do.
https://www.youtube.com/watch?v=0EdW3871-Qg
How about tool calling capability?
The video is showing the response of the model to the prompt "Write 200 things a CEO can do". This is not a very challenging prompt.
It gets to step step 144 where it starts responding increasingly in Chinese and then stops at 150 items (so fails to generate the full 200 items)
What is impressive about this? How often do you need incredibly long lists of things and are satisfied when supplied with only 3/4 of what you asked for?