Jump in the discussion.

No email address required.

I would like a fricking open model from the company with "open" in their name. Also I met a guy at my programming meetup who worked under Sam's mom (dermatology), apparently now she just fly's around in private jets lol.

If you haven't run deepseek locally it's literally 2 commands and pretty neat, Some other people in academic settings are using it to help find related papers to what they are working on. Really cool

Jump in the discussion.

No email address required.

you need 500GB of VRAM or something to actually run it properly tho

Jump in the discussion.

No email address required.

Neighbor there are smaller models, i run mine on a pi

Jump in the discussion.

No email address required.

Can you like link me to this rPI with 400GB of VRAM?


https://i.rdrama.net/images/1735397835BTbCkGwWb5B-VQ.webp

Jump in the discussion.

No email address required.

Jump in the discussion.

No email address required.

Literally everything in that article is false. It starts by talking about how bad ChatGPT is for the environment (not true) and how OpenAI recieved $500 billion of American taxpayer money (SoftBank is a private Japanese company!), then repeats the false $6 million dollar training claim, and the entire article is about how you can run DeepSeek r1 on a Raspberry Pi when in reality it's a finetune of an unrelated months old version of Qwen with 2% the parameters of actual r1

Jump in the discussion.

No email address required.

https://i.rdrama.net/images/1739238436pmsQfFB9dQ5NHQ.webp

Jump in the discussion.

No email address required.

Ty

btw I was being facetious and playing along with arseupkongo's comment


https://i.rdrama.net/images/1735397835BTbCkGwWb5B-VQ.webp

Jump in the discussion.

No email address required.

:chillguy:

Jump in the discussion.

No email address required.

>can you like me

erm, no?


dude bussy lmao

Jump in the discussion.

No email address required.

:marseysulk#: meanie


https://i.rdrama.net/images/1735397835BTbCkGwWb5B-VQ.webp

Jump in the discussion.

No email address required.

Not mean, just honest. You're pretty unlikeable.

Jump in the discussion.

No email address required.

:marseynotes:

Thank you for the feedback. I shall use it to improve.

Have a great day, Mr Goyim.


https://i.rdrama.net/images/1735397835BTbCkGwWb5B-VQ.webp

Jump in the discussion.

No email address required.

You're thinking of the largest models. LM Studio has a ton of smaller models that run on consumer hardware. I have 8GB :marseybee: RAM running Deepseek Coder v2 on LM Studio it just twerks !codecels

Jump in the discussion.

No email address required.

Are LLMs any good with higher end AMD GPUs? You get a ton more VRAM, but they're still not as great with video making software.

!codecels

Jump in the discussion.

No email address required.

Nothing's good with higher end AMD gpus because AMD is a garbage company run by r-slurs. AMD is never released a product worth buying that didn't go into a CPU socket

If u disagree ur just poor


dude bussy lmao

Jump in the discussion.

No email address required.

They're great for gaming! :marseypixel:

The 7900 GRE performs in between the 4070 Ti and 4070 Ti Super and used to cost $500. Now it's about $600+. :marseyitsover:

Jump in the discussion.

No email address required.

Reported by:

buying Ayymd

:#fellforitagainaward2:

Jump in the discussion.

No email address required.

Not sure it's still an issue, but AMD CUDA support sucked for a long time.

Jump in the discussion.

No email address required.

No, AMD is spectacularly bad with tensors, bus speed and memory throughput. AMD CPUs are preferred for ML workloads though.

LLMs are actually memory bound not compute bound. H series had much more dramatic improvements in memory throughout and bus throughput then GPU raw power. Nvidia won't let vendors do more than 24gb on consumer cards because they don't want them to compete with their DC cards.

The current rounds of model research is mostly on nano models that have much lower precision (so take much less ram to execute) and instead you build networks of them rather than just use larger models. Gemini 2 is very impressive, I'm working on some agentic systems right now with it.

Intel are actually making more progress than AMD on the GPU side for ML. This is one of the reasons Google have custom silicon for tensors, it dramatically reduces how much Nvidia they need to buy. Intel are making a similar play to try and claw back DC CPU from AMD.

Jump in the discussion.

No email address required.

CUDA is king, I don't even know if all the LLM stuff works on AMD

Jump in the discussion.

No email address required.

I can run deepseek R1 and other models easily, you have to just get the right one based on your :marseybee:RAM, I have 12GB so I can go up to 11gigs on the model size. work great

https://i.rdrama.net/images/1739231722GzSU1bqmiTnIvw.webp

also do you know why LM studio is now saying I can fully offload 8gig models? It didn't say that in the previous versions a while back

https://i.rdrama.net/images/1739231852TL5Et0oCgAVR2A.webp

Jump in the discussion.

No email address required.

well yeah the large one is the impressive one, there are tons of low vram models you could always run but they are mostly shit

Jump in the discussion.

No email address required.

you're shit :stoningshit:

Jump in the discussion.

No email address required.

Why are you so mean to him

Jump in the discussion.

No email address required.

I just wanna throw poop

Jump in the discussion.

No email address required.

Neeeerd! :nerd:

Jump in the discussion.

No email address required.

I eat shits like you for breakfast :marseyfluffyannoyed:

Jump in the discussion.

No email address required.

Oh I bet you eat shit for breakfast :mar#seyeatingbeans:

Jump in the discussion.

No email address required.

I met a guy at my programming meetup who worked under Sam's mom (dermatology), apparently now she just fly's around in private jets lol.

He transitioned?

Jump in the discussion.

No email address required.

In principle you can run r1 on your machine, but not if you have less than $10,000 worth of GPUs. Anything with less than 671B parameters is in reality just a finetuned version of Qwen or Llama; these models are nowhere near as good as actual r1

Jump in the discussion.

No email address required.

>programming meetup

Imagine the smell


https://i.rdrama.net/images/1735397835BTbCkGwWb5B-VQ.webp

Jump in the discussion.

No email address required.

not as bad as you think

Jump in the discussion.

No email address required.

tbh I actually do want to go one

But they're inner city and a bit hard for me to get to after work


https://i.rdrama.net/images/1735397835BTbCkGwWb5B-VQ.webp

Jump in the discussion.

No email address required.



Link copied to clipboard
Action successful!
Error, please refresh the page and try again.