GLM 5.2 is half of K3's cost on AI analysis, is faster tps, and less token burn. V4 flash is also above 50 (loosely getting more than half questions perfect) in AI index score and even much cheaper. This approach has much better cost curve potential, even if k3/fable are included in possible routing.
post
there doesn't seem to be anything here
1 comment in c/localllama@sh.itjust.works view that post
no comments (yet)