Dave

[Dave @HomeadePizzBest]

@Blackwellboy GLM5.3 is supposed to > Fable at that kind of thing.

[Dave @HomeadePizzBest]

@MiaAI_lab Dude.. get some sleep.

[Dave @HomeadePizzBest]

@MrCatid @jmurillocode you're saying that an external model is better than an internal one? Why?

[Dave @HomeadePizzBest]

@art_01x @MiaAI_lab Playing with this model it's clear to me it's strength is long horizon response - point being, who wants to wait 2 minutes for a low quality answer hehe.

[Dave @HomeadePizzBest]

@Tech2Wild hard to believe I read all that and the word 'cache' didn't come up once :) There's a cache for one and not the other, right?

[Dave @HomeadePizzBest]

@MiaAI_lab @mehmed_bazdar lol, unless you're a lawyer :)

[Dave @HomeadePizzBest]

@MiaAI_lab @VilmerRose glm5.3 is supposed to be good at long horizon stuff; how does that play out with this build?

[Dave @HomeadePizzBest]

@zhijianliu_ vllm version sniff sniff?

[Dave @HomeadePizzBest]

@MiaAI_lab Nice. If use dflash2 (+30%) and if requant to w4a4(+70%) I do wonder if I could hit 65 tok / s with no loss of intelligence.

[Dave @HomeadePizzBest]

@MiaAI_lab yup it's dethroned deepseek ... which I thought was it for me. It's building pac man for me now

[Dave @HomeadePizzBest]

@louszbd Just. Wow.

[Dave @HomeadePizzBest]

@Tech2Wild Yup - GLM5.3 is answering some amazing long horizon questions.

[Dave @HomeadePizzBest]

@ZixuanLi_ Running it locally and seeing a BIG jump in long horizon task processing. It takes time but comes out with better answers.

[Dave @HomeadePizzBest]

@Dave_Charland ok I'm early in these results but I have a battery of tests that I fire and so far ,
DS4: tools: check, cached queries: check, concurrency: check (DS0731 is near GOAT) but higher intelligence: 3/10.
GLM5.3 : tools:check, cache: check, concurrency : tbd, HI: 7 out of 10 .

[Dave @HomeadePizzBest]

@lucasfulks mtp-4 gave us little < 3 tokens per step and that pushed GLM to 25-34 tok/s.

[Dave @HomeadePizzBest]

@ayayalar and get another spark fro DS .. one spark is < 15-20 tok/s difference and quanted differently.

[Dave @HomeadePizzBest]

@ayayalar gotta be honest with you I would be bouncing all of those llm's against each other until they had something to say hehe.

[Dave @HomeadePizzBest]

@Tech2Wild glm is NAILING HLE questions HARD

[Dave @HomeadePizzBest]

@TechMDAI Yup 10 tok/s lol ..need 2