Excellent model ; curious about Reka 21B?
First - excellent model.
Downloading the source for "play time" right now.
I read your comments on the repo page, and in the community tabs "RE: reasoning".
This model:
https://huggingface.co/RekaAI/reka-flash-3
Seems to have a lot more compact reasoning/better reasoning than QwQ, Distill and others.
( I have been extensively testing/merging and messing with these)
Maybe tune it with story/rp ?
It is also 21B.
Also; quants of this model operate well above quants from other arch types too.
IE: IQ1_M actually works really well - but reasoning impaired.
However IQ2_S (augmented) works scary well.
REG: Q4/IQ4 are top of the class.
ADDED:
RekaAI also has 3 other models, 2B/7B/67B ; but not on HGF .
Asked them already if they will be adding/upload/allowing access to them.