How to run DeepSeek on your computer (100% private)
ฝัง
- เผยแพร่เมื่อ 5 ก.พ. 2025
- Get Unlimited DeepSeek R1 in Vectal: www.vectal.ai/
Wanna start a business with AI Agents? Go here: www.skool.com/...
I'M HIRING! If you want to help me build Vectal, apply here: forms.gle/NbLa...
Follow me on Instagram - / davidondrej1
Follow me on Twitter - x.com/DavidOnd...
Ollama - ollama.com/
Deepseek r1 - ollama.com/lib...
Anything LLM - anythingllm.com/
Subscribe if you're serious about AI.
Run Deepseek r1 locally and privately on your own computer.
www.vectal.ai
🚀 Developers, I’m hiring! Apply here: forms.gle/NbLaQNwAE6d9kkWs7
🚀 Wanna start a business with AI Agents? Go here: www.skool.com/new-society
it's important to have a free plan otherwise it is useless
As a non-American. I just care who is the cheapest. Sending data to the United States is as bad as sending it to China😂
Lmao yeah ok
Youre not wrong.
i dont care about "cheapest", free is more familiar in my mind
@@universe7779what will Chyna do with your data that you’re not comfortable with, that the us is that you’d be ok with?
Tell me that when China launches a war
Thanks!
When you run anything smaller than the 671b parameter model, you are *NOT* running R1. You are running a different base model that's been fine tuned on R1 reasoning outputs.
The base model for the 70b and 8b is llama 3.2
The base model for the 32b, 14b, 7b, and 1.5b are Qwen 2.5
Thank you for that! what would a system approximately cost to run it locally on a 671b parameter model?
Thats maybe true but it doesnt matter. What actually matters is the parameters and the template. So These are belongs to deepseek that means those models are also deepseek
@@Alex-si1bi I don't have an exact figure for you, but _a lot._ You'd need several hundred gigabytes of memory, I think I've seen about 800GB quoted, just to store the model. Then you need a little extra for the context window.
If you wanted to run it on CPU that might mean you buy an old server that has enough RAM slots to hold all those memory sticks, but it's going to be super slow. I saw a guy here on TH-cam run it on a server CPU and it took over an hour to answer a single prompt.
If you wanted to run it on GPUs you'll need a lot of them since the highest memory consumer card available right now is the RTX 4090 which has 24GB of VRAM. Or you can wait for the 5090 to be available with 32GB of VRAM.
For my setup I personally use a 3090 since it has the same memory as the 4090, but it's slower. I cannot run the R1 model though.
You can also try to run on an Apple Silicon computer which has unified memory allowing the GPU to access way more memory, but it's going to be pretty slow. Better than a CPU, but still slow.
At any rate, if you want to run the full model at any kind of speed you're looking at probably well over 10,000 USD just to start with.
@ It matters a lot. Most of a model's knowledge is gained during the initial training. The fine-tuning step is only teaching the the format of the output and is not very good at teaching the model new data. So if you use a poorly performing model as your base it's going to degrade performance significantly.
The reason they used Qwen and Llama is because they were the best performers at the time that also had licenses that allowed them to redistribute the models.
@@MrC0MPUT3R Thank you for your thorough response. Really appreciate it. Will digest this food for thought. Cheers!
bro please keep this going show me how to build a modular system so I can plug in any ai I want while maintaining our info privately
Ollama and LM Studio can do that.
Man, awesome videos recently, cheers! one thing, i'm struggling to see is how vectal is really groundbreakingly beneficial for $30/m???? feels like just a glorified windows notepad mix deepseek, but maybe im wrong
This is just a pitch disguised as a guide.
New here?
As an American I have no problem with china. The countries leaders are mad at each other. It has nothing to do with me. I’m going to use whatever best tools are out there.
You have to get past the first five minutes of babble
everybody knows that already . every "build anything" video path a way for his products . nothing about building something actually . dude is a genius
@@HoopsRaccoon What country leaders? Leaders in US change, opposed to China. And US does not seek to conquire the world. There is no doctrine about anything, again, opossed to China, which want's to spread communism across whole world and is already getting ready to attack Taiwan
Thank you for making that video David.
Unlike OpenAl, DeepSeek is open Al.
OpenAI = ClosedAI
DeepSeek = CensoredAI
@@2enZenHonestly all AI models should be censored to remove politics otherwise it will end up being biased no matter how hard you try.
@@2enZen😂try discuss Gaza with chatgpt. You will get a beautiful answer.
You are so censored that you dont know that openIA also have censorship, that is the level of unknownledge that you have.
Ask chatgpt "who is lunduke?"
@@Ateshtesh chatgpt is obviously not going to know every random youtuber especially those of small size. you can ask it who is "ateshtesh" and it wont know either, not because of censorship, but because it wasnt fed any data regarding you
How can I set it up for some text generation tasks instead the chat? I usually use paid API (Gemini/GPT) in my Jupyter notebook to summarize, extract text from a list of documents, so I am wondering how I can do it locally with DeepSeek.
Is there a dolphin deepseek on ollama?
Bro you look so much better in your facecam than in your thumbnail. My handsome brother please change the thumbnail face and do ur face justice. Great jawline.
I am running it on RPI% locally with PraisonAI multi agent framework, a bit slow but it works !
Savage: "deepseek was built for the same price as Sam Altman's car..."
Bro crushes openAI's capitalism with this statement.
exept it wasn´t. you people are unbelievable easy to manipulate and fed with b.s.
Who exactly told you that? Who exactly said that they used only 6 mi and old GPUs? Figure it out.
@AmazingSurrogats -- dont comment before watch the video. He clearly say that in the video.. that deepsek was built with very low budget.
If you want to suck OpenAI .. then do it. Dont get angry here.
I love to translate huge pdf files in multiple languages! 🥰
Reinforced learning is definitely the way to go with these AI models. Also never knew that Deep Seek was that popular that is caused a global outage. Wow.
Hey David, what was the name of the app for the iPhone?
After three times listening…it’s Apollo AI..right?
That's it
Very informative. Thanks for sharing
is there a way to fine tune deepseek with anything llm?
15:31 Bro really said if your l worried about
Ooooh no its going to China 😂😂😂
tldr for video. OpenAI vs Open AI?
Nice Job Dave ! 😁
Thanx for that video! Can you pls make a video how to integrate DeepSeek in n8n?
Can i deployed it via network and devices connected to same network can access it?
Great vid. Bro speaks at 3x tho - had to slow down the vid. Most other CCs I have to speed them up 😂
Saying Sam Altman has some ego with his OpenAi model when you literally say every video that Vectal is the best tasks tool 😂
Who cares man, everyone has ego . You do too
It’s unbelievable!USA $500 billion AI competing against CHINA $6 million AI…wooo not bad at all. It just show how good brilliant are those young Chinese engineers, mathematicians & scientists are in CHINA 🇨🇳 today..DEEPSEEK RI is just the beginning! 😊😊😊
:-)) Really? Tell me exactly who told you that? Who told you that they only used 6 million and a few old GPUs? When you figured it out, you ll realize that it´s not really wise to repeat everything somebody spills out.
Will the new rtx 5080 gpu be very good for running locally`?
If you offer a service, and it's paid, it's not really for everyone.
Hi, I have a Qualcomm NPU on a Windows ARM, can someone help me run it on an NPU?
NPU is very underperforming with low memory and using shared system memory - it is good for small models like Whisper speech to text. I made comparison and even iPhone is more powerful than NPU (iPhone can use WebGPU in Safari which enables you to use Huggingface transformers and ONNX Runtime to load AI model to iPhone using WebGPU and it will use iPhone GPU to run the model, not CPU, so iPhone GPU is much faster than NPU - you can ask AI and verify it)
What are the minimum system requirements for running Deepseek AI on my personal computer?
Thats not Deepseek. THE true DeepSeek is the 671b MoE model. Everything else is either Qwen or Llama3 finetuned with the Deepseek Dataset.
What do you think about LM Studio?
Locally means loosing coding capabilities that is NOT the most wanted, since mostly people care for the coding assistant help!! Isn't there any configuration with CodeGate for Cursor AI or VSCode / Windsurf etc?
Watch fireship beyond's video about it.
@@Bytefy link please
@@Bytefy I haven't even heard about what you are saying! So it would be nice to navigate us please with a link!!
Bro do you ever come up with any solutions on your own?
we're not all scientist are we?
You work for 42yrs to have $2m in your retirement, Meanwhile some people are putting just $20k in a meme coin for just few months and now they are multi millionaires I pray that anyone who reads this will be successful in life
Well explain thank you for bringing up this video Financial education is indeed required for more than 80% of the society in the country as very few are literate on the subject. The value of the US🇺🇲 dollar is declining due to inflation, but it is increasing in comparison to other currencies and commodities such as gold and real estate. I'm worried that rising inflation will cause my R$550k in my retirement funds to lose value, But with the help of Mrs Sonia I hit $220k this week from my investment of $45k, I am truly grateful for all the knowledge and nuggets you have given me over the past few months.
I have been seeing so many recommendations about Mrs Sonia , she must be really good.
Luck you you too buddy!
Every cell phone provider has had data leaks. Your data is already exposed. Carry-on.
Binance infinity ETH bug right now
I just made a video to show that_
add "--verbose" without quotes at the end of the prompt to get also some nice statistics at the end. » ollama run deepseek-r1:8b-llama-distill-q8_0 --verbose
does the thinking portion of this LLM count towards the token costs, or is it just the solution portion?
Not sure, but I read OpenAI O1 simulates the reasoning by running different prompt queries in parallel so each reasoning paragraph is a new query with new token costs and new context window. So your initial prompt is not influenced by reasoning portion! It is like using 5 AI models during the reasoning and then it asks to create a summary out of it and that's the final solution portion you get.
Disconnect from the internet also in case there is a malware
Amount of money that OpenAI makes: -1billion USD per year
number of advancements OpenAI provided to the world and to science: A total of 0
Once setup and running on a decent PC how do I keep it all private with nothing going out of my LAN and how can other people in the house with a browser access my setup?
decent PC is not enough. don't get too hyped about this offline ai solution. the reality is it's too expensive to operate.
nothing going out of LAN is the default setings when you run model yourself and then you use WebUI for others and it is not going out of LAN because you dont have public IP address, but for DeepSeek you would have to buy 8XH100 which would cost you 350 000 USD. And then you have it FOR FREE muhehehehe....sell that house and Audi or BMW and "hold my pepsi"
for the 70B, is the 43GB for RAM or storage?? 🤔
43gb storage , and 128gb+ vram , that means 4x5900rtx nvidia cards
How do I turn off the automatic translation?
In vectal would I get full access to DeepSeek? I am currently running the 32 billion version locally on my computer but it’s very slow. I am processing very large files and I find that most online services truncate the output.
What kind of MacBook you got? That you able to run 70b??
Year 2080: the community still waiting the response
Seems that M2, M3 chips are able to run the model because they have shared memory between gpu and cpu. Search for UMA
Actually where are those LLM install on the hard drive and how can we see the source code ?
Use deepseek r1 with cline locally and test it please
I'm new to everything. I need to know if I install Deepseek locally, will it autopdate with new responses depending of the NEW NEWS via internet? For example if tommorrow Donald Trump dies, and I installed deepseek today locally (the biggest version 70gb) will the AI know tommorrow that Trump is dead if I ask him?
I got "Error: model requires more system memory (16.4 GiB) than is available (11.1 GiB)"
What do I do?
you would enable offloading to CPU / RAM or using quantized model from HuggingFace. Instead of Floating Point 32 or BFLOAT16, your model will use INT8 or INT4. This will loose precision, similar to converting PNG to JPEG but you can find middle level compromise where you get still good quality results - I think vLLM for serving models or SGLang has the best options to compute KV cache and offloading individual AI model layers to CPU on the fly, when needed. This way you can run most of models on consumer GPU.
but it is hardware intensive right? tried it and it uses about 80% of my cpu
Is DeepSeek R1 multi-modal?
No 😔
Thanks for ditching that cringe, Alex Hormozi nose strip thingy.
6:25 im always afraid of terminals, even though i do a little bit of programming myself, sometimes make mods for games etc.
does it work on windows and...i did not understand anything
can you run these local models with local n8n?
Can someone explain to me how to use DS on Vectal?
how to make text to speech in deepseek? thak you.
ask deepseek
so you are telling me that browser version is not safe?
if you're trying to invent something and you don't want your world changing new ideas to be read by someone else, then no, browser version is unsafe. but if you're only going to be using it for academic purposes then yes, it's safe.
How can I disable the reasoning step () in DeepSeek R1-14B model? I want it to give direct answers without the explanation. thnks for video!
I think R1 do it always same like O1. If you dont want it, then you would use DeepSeek_V3 model. Or there might be some system prompt override.
Unfortunately even though I bought a new iMac with M3 chip they use integrated GPUs making running models almost impossibly slow.
It’s a lack of total ram. The m4 have 100g of shared ram
Can i run this offline?
Advice: You are very enthusiastic and speak quickly, but your pronunciation of words can be unclear since it's not your native language. This can be confusing for your followers, who come from various linguistic backgrounds on social media. It's important to control your pace and speak naturally without overdoing your enthusiasm.
My greetings to you! Also, please remember that using bots on your account can be very annoying and is not a good idea, as followers can easily detect it.
This is definitely Skynet
we're getting there 😂
Great thanks!
How tf you watched 20 m video in one minute
I was the 500th thumbs up! 😇🤠
Glad the bruise on your nose has finally healed
What can I run on RTX 3060 with 16 GB ram?
Hi! I suggest you use the distilled Qwen 7B or Llama 8B. You could also try the distilled Qwen 14B but it will be notably slower.
Here we gooooo 1 token per second 🤣
DeepSeek❤ China💥
Tell that to brainwashed india
@Siffat22 India: only no.1 in Over-Population in world in all other fields India Completely 0 loser👎👎😂😂
India is lagging behind in all other fields except population.
@Siffat22 INDIA: only no.1 in Over-Population in world in all other fields INDIA Completely loooserrrr........
INDIA is lagging behind in all other fields except population.
Why are you always so excited? I thought I was looking at 1.5 x
but who told you 800k people are gonna get benifited with ai?
Why does YT erase my post notifying David I sent a message on IG ?? I don’t get it
Ask chat GPT to summarize DS privacy policy
Buy the dip Nvidia for sure nothings changed about there product offer
Is this censored?
ask chatgpt "who is lunduke" and then put in here the answer.
Why doesn't open AI change its name to Close AI😂
Diky z jicina budem ta sledovat spolu s 799000 dalsich
not worth it, any model that shows its thought process soon after using you will realize a lot of time is gonna be wasted.
This channel is still on the border between sensationalism and ridiculousness... :D
Dude, are you running an auction or explaining a complex subject? Just speaking fast does not make your Russian accent disappear.
It's better to have my data with china than to have it with my own govt or companies from my own country. Just think who can have reach you first.
Cool
Is Ondrej AI generated?
I would rather send my information to the Chinese company because it’s free or a gift from the Chinese. I don’t trust my American government anyway and it’s not free.
Ohhhh……. It’s soooo not private 😂
It was good, but you are to fast.
i did this, and the one you run locally is super bad compared to the one online. I mean ATROCIOUS!
😭😭bro vectal is paid
hi
Pls show us cursor with r1
First comment ❤
😇
You want to use a model from China but you keep dissing China... seems like you're letting some stereotypes hold yourself down =D
Why are Americans so worried sending their data to China, I mean do they trust their government any better 😂😂
It's not R1 😂
This guy rates himself $500 hour and he is full of ego. Hope someone teaches him a lesson by making him taste his own ego
Bruh not only chinese AI models are collecting data. Every single tech company has it's user's data no matter the country. How did Meta train it's image generator and LLM?
Screaming like a parrot at everything someone chews up is not exactly a sign of special intelligence. What exactly is opensource?
And who exactly brag about using only a few millions and a few old GPUs? The "start up" in china with 6 employees? Really?