ReAnRe

.

[ReAnRe @RevAndrei]

@JASONMCNAB @MiaAI_lab @NVIDIAAI @NousResearch this is exactly what I run now. Seems solid so far.

[ReAnRe @RevAndrei]

@MiguelCamba this is sick

[ReAnRe @RevAndrei]

@0xWhiteMage @Blackwellboy @MiaAI_lab I am afraid you need one more spark - the math simply does not add, you will risk to lose too much quality. Still... from my tests Qwen 3.8 flash next will do fine. Btw... what is your use case scenario?

[ReAnRe @RevAndrei]

@0xWhiteMage @Blackwellboy @MiaAI_lab It is the first time when someone called me a bot. I am not sure if I should feel honored or angry. Anyway, just let me know if you want me to clarify something.

[ReAnRe @RevAndrei]

@0xWhiteMage @Blackwellboy @MiaAI_lab Who is Mia... does not actually matter. If it is a man or a woman, or a company.. I could not care less. What is important for now is the common goal. the fact that they are making money out of it is good, this means that they are not controlled by an immportant big player.

[ReAnRe @RevAndrei]

@MartinSzerment But do we actually know the real numbers (parameters) behind the closed models? Or how efficient is the architecture?

[ReAnRe @RevAndrei]

@TechMDAI The amount of money that nVidia is doing by selling DGX Sparks is rather small. What they are actually doing is totally different - they are protecting the datacenter investment of other companies, to basically prevent wide spread adoption of local AI inference.

[ReAnRe @RevAndrei]

@MiaAI_lab @StefanMaier Each and every one of us who is dealing with local AI, who is tinkering with variour models, who contributes in one way or another to the adoption of local AI has a bit of MIA dna. No matter the name, nationality or preferred model.

[ReAnRe @RevAndrei]

@MiaAI_lab this will only encourage development of model that would fit in 2 lower memory sparks, which means that we are going to have even more efficient models in the future

[ReAnRe @RevAndrei]

@Mr_WestCoast * don't get it. The person simply wanted to stay anonymous.

[ReAnRe @RevAndrei]

@MiaAI_lab only if they increase the bandwidth as well

[ReAnRe @RevAndrei]

@MiaAI_lab I would have never expected in my wildest dreams that I would get this level of intelligence and speed on my 2 sparks. Every night the agents are working like crazy on my php projects - one more month and those toys are going to justify the return of investment.

[ReAnRe @RevAndrei]

@jayleaton So basically you fixed the issues with the quality, while improving speed. Great!

[ReAnRe @RevAndrei]

@ashxhart There were many issues with tensorfold, which is normal, but the project is growing beautifully.

[ReAnRe @RevAndrei]

@landontgreen I just made some bench with Tool Eval - 87/100, really weak, because it does not seem to have structured output. Now... the good thing is that TensorFold 0.5.0 has support for this. If I am not wrong this recipe woud be incredible if the quality will be fixed. By far the fastest

[ReAnRe @RevAndrei]

@minviable_org @landontgreen I read now that the context problem is solved, will try it again at the end of the week.

[ReAnRe @RevAndrei]

@DearS_o_n It depends in what country do you leave. I felt like this in Germany, but when you will go in Eastern Europe, people will be different, especially the older generation.

[ReAnRe @RevAndrei]

@ashxhart @TheAhmadOsman unfortunately I could not put enough context in GLM 5.3 using tensorfold. maybe I am doing something wrong, but 64k of context is not usable.