do not rely on internet AI models but instead put it in your own computer to use it offline

use it on offline linux to maximum effects, windows 10 doesnt even work properly if disconnected from the net?

https://ollama.com/

download for example Chinese deepseek1 onto your own computer, 5GB model, 20GB, model, even 400GB model its all there

400GB model is a bit problematic to get if you have 1MB/s internet speeds, it will take 4.5 days of constant downloading 24/7

I dont even know if the site is able to continue download if its cut of suddenly, from where it was previously, they dont provide torrent type of a download

But all the other models are easy even for 1MB/s speed

A reminder that I didnt experience a connection being cut when downloading 20GB and dont know what would have happened if there was such occurence. I have roughly aobut 2MB/s connection.

>shithub
>no torrents
no thanks

>>509899945
I know it sucks to not have a torrent

>>509899205 (OP)
4.5 days of constant downloading? Shit... Yall would've hated 56k V90 modems...

But is it as easy running a .exe?
I have gigabit internet*, but I'm also a giga retard.

*even though i live in the ass end of nowhere paying the the cheapest plan available to me

Using locally hosted AI? That’d be like taking advice from a retard. The whole appeal of AI is amount of knowledge, and no computer can compete with a fucking datacenter the size of a small city.

>>509899205 (OP)
nigga, you're not gonna run gpt-tier LLMs on your 8gb of VRAM. you can run async storytellers, that's about it.

>>509900336
I think you need 12GB free for even the smallest model. Under that it just doesn't do anything, other than 'What?' maybe.

>>509900268
you can get good enough answers completely free and offline, the issue is that it takes a long time with your own hardware, compared to their warehouses. but your initial point is false.
you can also feed it any knowledge you want but the issue is compute, not whether or not its connected to the internet or not.
the deepseek models that OP is talking about are very good.

>>509900268
Seems like you don’t know how LLMs work. The knowledge is condensed into the model. The 400gb model would have just as much knowledge as the one hosted in said data center. Even so, the smaller models are still pretty good

>>509900487
more or less, small models can be quantized but they'll be slow. phi-3 or tinyllama can easily be ran at 8gb and you'll get a lot better than just 'what?'

>>509899205 (OP)
sanna, please stop drunk posting and at least provide something normie-proof like
https://pinokio.co/

>>509900660
What?

>>509899205 (OP)
Ollama is fake and gay Jewish shit.
Use llama.cpp and huggingface directly. Ollama is just a wrapper made by yids that actively makes everything shittier and harder to use.

>>509899205 (OP)
That chink ai is bad. Even worse than gemini. ChatGPT and grok are vastly superior than the rest.
Grok 4 will turn everything upside down. It was released today.

>>509900593
Diminishing returns hit hard and the 400gb model isn't really better than the 40gb one. Really it matters only for knowing specific trivia facts.

>>509900724
All AI is slop and stop selling the Jew shit here. There is no difference between chinkshit AI and SOTA Jewish AI, they all hallucinate like crazy and can't into basic arithmetic and logic.

>>509900251
>running as an .exe

on linux it is, just type on a terminal
./thenameofyourlanguagemodelfile and press enter

(you just make sure linux file permissions have been set for this file to be executable)

>>509900833
yeah these people are retards, all the models are good the companies are just running a bunch of background functionality over and over again to provide you with up to date context information. as well as remembering everything youve ever said to them as context.
thats why theyre "better" not because of the model but the infrastructure.

>>509900524
>>509900593
Yeah, I’m talking about computation, not storage. Answering Wikipedia tier questions is different than producing videos or creating complex code

>>509900724
It hasn’t been unveiled yet, won’t be until tonight at 8pm pacific time. The benchmarks are very good if they’re true

llamacpp is not difficult to use and you can use many more LLMs much easier and faster

>>509900972
The trick is having everything stored on your own computer and being able to execute them offline.

Whatever the model is that you are using.

>>509900915
Is it accessible only througfh an X service?

>>509900914
The reason big companies need large data centers is for training their models and for making them usable at scale. Making a pre-trained model available to one user doesn't require a data center, although you might need a high-end GPU setup for generating videos.

>>509901099
nah you can access it through their website directly without any twitter account. but grok has been pretty shit and always light years behind on the only benchmark that matters: https://simple-bench.com/

And how exactly do you guys plan on squeezing a 400gb model onto 8gb of VRAM? Maybe if you're lucky you'll get speeds of 1 token a week (if your computer doesn't shart out due to lacking ridiculous amounts of useless RAM). Just ask the gooners what a good model is if you're a VRAMlet.

>>509901078
That is dhat makes using llamacpp over ollama such a better option. With llamacpp you have the direct files instead of what ollama decides to allow to be downloaded via ollama

Also all the best uncensored LLMs are not available via ollama. You can download the gguf files and use them directly via llamacpp

You need to go a few steps further. Llamacpp is where it is all at

>>509899205 (OP)
Just press alt+F4 to get a faster download

>>509899205 (OP)
llms are useless fucking garbage for actual brainlets
kys

>>509899205 (OP)
r1 has neither the feature set nor accuracy i need for AI to support me effectivly in my daily workflow

t. rtx5080 owner that tried selfhosting

>>509900251
rm -rf /

is linux superpower to enter eZ mode