14GB, that is what being said about!
Bigger, but speed is almost identical to the smaller variant. It is smarter than 12gb one, but not to a new degree really, it nails dialectic nuances, seems like good at coding(tested on html game - APEX made almost the same one with first try, so seems no different, but runs faster, and Vision is usable with Cerebellum, because it somehow lowers RAM consumption after being loaded by 1 GB mostly all of the time) Waiting for Heretic one, if it will come at all.
No update for it? came out? not as good as it was expected?
That was a link to the update. Both versions are there. V1 and v2 heretic. Unless I'm misunderstanding?
That was a link to the update. Both versions are there. V1 and v2 heretic. Unless I'm misunderstanding?
Update of how it smarter and what changed(in tests and overall quant info). Also, would you try to do Gemma 26b next? or you stopping for now? Because today downloaded again Cerebellum gemma 26b and it is very good, but hallucinates in tasks where 4 XS dont, wonder if bigger version would beat 4 XS one, or is it the one im stuck with(but it is bery capable and super smart, yet so lazy if you ask too much)
That was a link to the update. Both versions are there. V1 and v2 heretic. Unless I'm misunderstanding?
Update of how it smarter and what changed(in tests and overall quant info). Also, would you try to do Gemma 26b next? or you stopping for now? Because today downloaded again Cerebellum gemma 26b and it is very good, but hallucinates in tasks where 4 XS dont, wonder if bigger version would beat 4 XS one, or is it the one im stuck with(but it is bery capable and super smart, yet so lazy if you ask too much)
Definitely not stopping, my brain just jumps around, so it might feel like I'm stopping, but I'm either working my regular 40hr a week job, spending time with the family, or working on anything my brain thinks is a good idea. Sorry the push is slow. But you testing the models further than I could, definitely helps me know where to target. So don't worry not giving up, just squashed on time, with a brain that struggles to stay on track.
Any other issues with the Gemma 4? It would help me narrow down where to go.
That was a link to the update. Both versions are there. V1 and v2 heretic. Unless I'm misunderstanding?
Update of how it smarter and what changed(in tests and overall quant info). Also, would you try to do Gemma 26b next? or you stopping for now? Because today downloaded again Cerebellum gemma 26b and it is very good, but hallucinates in tasks where 4 XS dont, wonder if bigger version would beat 4 XS one, or is it the one im stuck with(but it is bery capable and super smart, yet so lazy if you ask too much)
Definitely not stopping, my brain just jumps around, so it might feel like I'm stopping, but I'm either working my regular 40hr a week job, spending time with the family, or working on anything my brain thinks is a good idea. Sorry the push is slow. But you testing the models further than I could, definitely helps me know where to target. So don't worry not giving up, just squashed on time, with a brain that struggles to stay on track.
Any other issues with the Gemma 4? It would help me narrow down where to go.
Understandable. With Gemma 4, well it hallucinates quite frequently if asked something a bit more specific. Yesterday I was forced to use cloud Qwen model to compare results from two Gemma versions, in case I missed something. I asked for a few another tests and in result 4XS - won. The thing is it knows pretty much and understands very narrow traditional regional aspects, while maintaining clear language without hallucinating, creates better structure of answer(and even expands it(while Cerebellum one(which is smaller) I thought that it exceeded in creating new things, but it seemed on first generation, it was just lucky one, because other times, it was not as great). With those tests I understood that model is very capable and for the size! it is only 14gb and it knows so much, it is incredible. Sowhile Cerebellum is superior for its size, but adding two gigs on top makes it much smarter. Also first time I tried Gemma 4 w6b it was 3KP(higher than KM), I thought well it is pretty reasonable quant, but foreign languages was totally broken and it was much, not it was unusable despite weighing more than Cerebellum(it was 13gb or something).