[dofm]

I really had high hopes for the larger Ternary Bonsai and it feels like there is scope to improve, but I get the sense (albeit a naïve, probably not fully informed sense) that improvement can perhaps only come by training directly into ternary.

[kamranjon]

I’ve actually been really impressed with the 27b model they recently released - amazing performance approaching 40 tok/s on m4 max and I didn’t run into any quality issues in the small set of tasks I tried. Haven’t gone full coding with it yet but suspect it’s better than say a 9b or 12b model.

[dofm]

I ran into some issues that are more extreme versions of Qwen's thinking loops while reasoning. It is great at the small puzzles I set for it but it did some frankly insane things on a PHP coding task I set it.

It also had some issues that might be parsing/chat template stuff, tool calling oddities. I will try it again, I did try it pretty much the day it shipped and it's possible there are more improvements in their llama.cpp fork since.

It would be churlish to be overcritical, mind you — the PrismML ternary stuff is an advance — but it feels like it should be applied at training. I figure we will see that, somewhere, quite soon.

Did you try the BottleCap ThinkingCap Qwen post-train with the reduced thinking overhead?

[kamranjon]

I haven't spun up thinkingcap yet but I'm aware of it and am intending to try it out soon. How did you find it?

[dofm]

Not tested much but it is not noticeably worse than the underlying Qwen 3.6 27B in Q4_K_M, which in my experience is kind of a first for a Qwen fine-tune of this nature. They are almost always worse.

I think it does use fewer tokens while reasoning, which is potentially useful. I need to do more testing, because any performance advantage over the 27B is useful for me on an M1 Max.

[embedding-shape]

> suspect it’s better than say a 9b or 12b model

Whaaat, a 27b model might be better than 9b or 12b model? What would make you do such an outrageous claim?

[kamranjon]

Sorry I should have clarified - I meant that a ternary 27b model would outperform a non-quantized or 8 bit quantized 9 or 12b model - which it is generally close to (or much smaller than) in size. So yeah the comparison I was trying to make was between models of equivalent size or models that could run on similarly sized hardware.

[embedding-shape]

Ah yeah, that makes a ton more sense :) I mean, what you said earlier also makes sense but was too obvious, now it makes sufficient sense, thanks for explaining :)

[avadodin]

All you need is Ternary Aware Training and for AI researchers to come up with a backronym for TIT.

[woadwarrior01]

Wouldn’t that be TAT?