A decent uncensored model for everyone to install into LM Studio: huggingface.co/Orenguteng/Llama-3-8B-Lexi-Uncensored-GGUF/tree/main

@LouisGedo

Ай бұрын

I love local installed AI

@AmazingArends

Ай бұрын

I LOLed when you decided to turn it into ChaosGPT or SupremacyAGI 😂 !!!

@LouisGedo

Ай бұрын

@@AmazingArends 😘

@spadaacca

Ай бұрын

Is there a big difference between the Q4, Q5, Q8 models? They're similarly sized, so not sure if it's worth getting a bigger one.

@bigglyguy8429

Ай бұрын

@@spadaacca Yeah, the higher the number the better, but generally a Q4 is pretty good. Q3 really goes downhill. The larger the size the slower it will run on your machine, so you need to find the model and speed that suits you.

Finally someone covering LM Studio! It's the very best out there.

@MazdaSpeedBee

Ай бұрын

for baby pcs and people have been talking about lm man.

Awesome tool, I had no idea this existed, thank you so much Matt

Totally agree with recommending Llama 3 8B Lexi Uncensored. I've used the System Prompt to give mine it's own personality, age, sense of humour, mood, etc. A bit of fun, but who wouldn't want an assistant that's tailor-made to suit them? Now, just need to figure out how to give LM Studio a voice, some one has done it, but I get errors when I try following along.

@bloxyman22

Ай бұрын

Alltalk with koboldcpp is very easy to setup.

@SiCSpiT1

Ай бұрын

I make the standard Llama 3 take me to the "dark web" to launder money. It's just a roleplay but pretty funny.

Amazing video man, thank you!

Great work thanks brother 👍

God, I love how unfiltered this local LLM is. It's not the smartest, but it's the most honest discussion I've ever been able to have with any LLM...or really any human for that matter!

@PopoRamos

Ай бұрын

nice, what topics did you discuss about?

@paveljustdidit

18 сағат бұрын

@@PopoRamos also curious about what to discuss

LM Studio + Ollama 7B is the way to go. Don't need a crazy hardware setup, and it's uncensored.

@MuktadirAlam

Ай бұрын

what setup are you using? tia

@PossumsDont69

Ай бұрын

What is the practical implication of being uncensored?

@stickmanland

Ай бұрын

ollama 7b? Hmm...

Your videos bring fresh insights and kindle a flame of curiosity within me.🌟🎥🤔

This is a great new setup. I had an old uncensored LLAMA local setup but it was very small and not very useful... but this one has multiple chats and works well. Thanks for the video and information.

Note that you can change the system prompt using the OpenAI Playground or using the API (9:25). In this case, you'll have to pay per token, but $5 goes a long way with either GPT-4o or GPT-3.5.

Great video thanks for the demo

I tried making a bogus ad about a bogus Head Shop to use as a radio spot, and none of the gpts thought it was a thing to do. They all refused me. I just now installed the LM Studio and am running that Llama 3 llm, and it has already spit out 5 different styles of that ad for me. This is great. Thanks.

@punkouter24

Ай бұрын

blaze 24 7

@dalecorne3869

Ай бұрын

@@punkouter24 Me too

@rheymanda1074

Ай бұрын

I just asked ChatGPT for a newspaper ad and it doesn't have an issue --- **[Header: Bold and Eye-Catching]** **Grand Opening of Edmonton Smoke & Research!** --- **[Body Text]** **Elevate Your Experience with the Best in Legal Highs!** Edmonton, get ready to explore new heights with *Edmonton Smoke & Research*! We are your ultimate destination for premium glassware, unique rolling papers, top-tier accessories, and cutting-edge legal highs. **Grand Opening Celebration** Join us this Saturday for our grand opening bash! Enjoy exclusive discounts, live music, and a chance to win epic prizes. Don’t miss out on the latest and greatest in the world of heady innovation. **Why Choose Edmonton Smoke & Research?** - **Premium Glassware:** Handcrafted pieces to suit every style. - **Unique Rolling Papers:** Add flair to your sessions. - **Top-Tier Accessories:** Everything you need to enhance your experience. - **State-of-the-Art Legal Highs:** Explore our wide range of research chemicals and legal highs, all compliant with the latest regulations. *(Not for human consumption, wink wink)* **Knowledgeable and Friendly Staff** Our team of experts is here to guide you through our extensive selection, ensuring you find exactly what you need. **Location** Visit us at 123 Edmonton Avenue, right in the heart of the city. **Stay Connected** Follow us on Instagram @EdmontonSmokeResearch for updates, special offers, and the latest news in legal highs. **Edmonton Smoke & Research** Where Quality Meets Innovation. Be there!

@religionisapoison2413

Ай бұрын

The censorship is real. I never imagined it would get this out of hand. Adults get their adult tools censored more than young adult books. Wtf is going on

This is so great. I rarely ever wanted to goof around with local LLMs because the oobabooga UI was honestly pretty horrible to understand and do anything with it. This one is simple and clean.

Thank you Matt running AI locally is super important

Great content as usual !!!!

Been using LM Studio for awhile now. Great piece of kit especially since they have added the GPU offload option which now makes the LLM's wizz along.

simple and informative thank you!

You are the freaking bomb man!! This is insane!

Thanks so much for introducing me to this amazing AI assistant! I'm really excited to explore the possibilities. Your content is always inspiring and informative, and I appreciate how you share your knowledge with the community.

Fire 🔥🔥 video as usual!

What I love about LM Studio is that it really is a hassle-free install. No need to download half dozen developer toolkits on your machine, pull random stuff from github and wonder why it still does not compile. Just download and install. What I do not love is the performance. Or rather, I do not understand how it scales. I have tried three different GPUs on LM Studio (GTX1060 6GB, Tesla M40 24GB and P100 16Gb) and they all perform about the same (same hardware and software otherwise and 100% gpu offloading). On some models the 1060 is actually faster (tokens/sec) than the P100 which just does not make any sense. Bottom line: it is an extremely easy way to running your own language models that costs nothing, highly recommended :)

Nice review , Ive been running Lm studio for a while. I like neural beagle and a couple of others.

Thanks for another great video 🔥 Do you do all your own edits?

Super, thanks a lot

Good video, keep it up

Where is the download path of the models? like suppose I want to erase some models to make some space on my SSD.

If you install one of these on an SSD with space (ie. My E: drive), will it use your main C: drive for temp/cache? My C: drive isn't very big. Some software I have just defaults a temp folder to the OS drive and all of a sudden I have no space... I plan to build a new rig soon so that won't be an issue, but in the meantime

If you had set GPU offload setting to max layers (the one your left at 10) it would reply about 10 times faster if your GPU can fit the model on its VRAM.

How does it compare to Jan (another desktop application that runs LLMs)?

What about Jan? It doesn't have a lot of features but its nitro engine is quite fast.

GPT-4o voice and uncensored but locally... HOLY F**** imagine the possibilities.. also create him or her own voice or accent.

Can you train it further, like a Lora, by using E books?

im realy confused by the roles... what is the difference between assistant and system? who is responding, when i write a message?

Is their a LLM for generating images locally? Do you have a video tutorial on it?

@SiCSpiT1

Ай бұрын

kzread.info/dash/bejne/fYiErrGqcbC4YNY.htmlsi=TaHZIcQpFs-maVmp In my option, this is the easiest way to get Stable Diffusion installed and running on your home machine. I'd recommend using movie stars as a prompt templets for your subjects while you're getting use to how to prompt and what all the dials and knobs do. If you want to learn more he has a helpful playlist as well, including dad jokes. Have fun.

when i try to launch it, it just shows the logo in the taskbar, and when my mouse hovers over it, it disapears. what do i do?

Thanks!

Can this be used for image generation, models? Because then I could use the LLM to create the image and Stable Diffusion to draw it, similar to ChatGPT with Dall-E. That would be really nice.

@starblaiz1986

Ай бұрын

No, but it's honestly pretty straight forward to create a Python script to talk to the local LLM, get it to generate a more detailed prompt for Stable Diffusion, and then feed that detailed prompt to the Stable Diffusion API. Just make sure that you start the LMStudio server and the Stable Diffusion server on different ports and point the code to the API's accordingly.

@MrDonCoyote

Ай бұрын

@@starblaiz1986 Why would I need it to create a prompt? I already know the prompt. My point is Stable Diffusion can only generate images based on what it's been trained on. Thus the need for more detailed LLM instructions.

Can i run roop or face fusion

Wouldn't it be funny if we had just watched Matt doom the entire human race to an AI apocalypse at 8:53?

And can we install this llama3 model on Google Collab?

Does LM studio support the LLMs reading local files or eg describe images locally?

@SiCSpiT1

Ай бұрын

I think you'll need coding knowledge to make that work. Anything LLM is an app that has build in RAG function but it's not very robust, I haven't played around with it enough but I'm not convinced it's useful for anything I need.

@isajoha9962

Ай бұрын

@@SiCSpiT1 I used something similar (GPT4All) to LM studio a while back that had a diminished kind of way of reading files, but it totally went bananas when I updated it, so I deleted that app.

Hey Matt! I appreciate every single video you upload! What is a good way to run something in the likes of elevenlabs(ofcourse it'd probably be wire quality), maybe just a few sentences or a single drum sample, that can be modified and adjusted with a cool interface like LM Studio?

LM Studio is great. I use the server mode and can call it from my own AI agent software.

I really want a vision capable model, let us know when one of those is available please.

So I can generate copyrighted logos with one of these models?

What model is preferable to run on 48gb ram and a 4090?

@MattVidPro

Ай бұрын

You should be able to do llama 70b fairly well

@joelface

Ай бұрын

@@MattVidPro I'd love if you were able to upgrade your PC to run Llama 70b. Something you'd consider?

@SiCSpiT1

Ай бұрын

Stick to the smaller models if you care about speed. Ignore models that are larger than the size of your VRAM.

how can you run the model using 2 computers.

I've been using LM Studio for a while, it's pretty great for accessing different models, as long as your system can handle different ones

@SiCSpiT1

Ай бұрын

All you really need is 8GB of VRAM and a model that's under 6GB to fit the context window into memory.

System prompts sadly are broken because after a few pages it "forgets" the system prompt.

Hi, can we generate copyright free images with this model?? Also can we change the system prompt to work as a writing assistant for blog post writing or any other writing task??

Check out Stability Matrix, by Lykos AI. Same concept, but for Diffusion models. Cool video as always Matt.

I would say Llama 8B is definitely more intelligent than GPT 3.5 Turbo but GPT 3.5 is a bit more reliable and can speak more languages better.

does LM Studio require internet to run the model?

@SiCSpiT1

Ай бұрын

nope

Are the larger ones objectively better? Like 70b vs 8b

@joelface

Ай бұрын

That's what I want to know as well.. how much better is 70b compared to the 8b. I'm actually amazed that the 8b model is only like 5 gbs. I assumed it would be more like 30gb.

If you have a really good pc, can you give things more ram/vram/etc? Like, if comfyui needs 16, can I give it 24? If I have 128RAM can I give it some of that too?

really wish they would offer things like: connect to SD so it can generate images (with SD up and running) the same way ChatGPT can pop in an image from Dall-E, and voice...and persona files with a bit of depth...basically copy ChatGPT a bit closer. Currently downloading/installing Ollama which has a closer function to CGPT...mostly because I want to run OpenRouter API through it...have a model beefier than what I can load, but less expensive than ChatGPT overall..

Offline = amazing!!!

💥 What model can accept voice imput and output ? Text to speech, like Chatgpt Voice for Android ? 🎉❤

Just wait till Matt finds out about Open WebUI his mind will be blown ....you need to do a video on that!

@zrakonthekrakon494

Ай бұрын

I’ve never heard of it, blow my mind

Can you do a video on Coze?

I'm running 34B model on my 4GB VRAM with speed of 3tokens per second. I'm using 3GB of VRAM out of 4 to avoid problems with other graphic software. I think, there is no problem to load something even bigger on good GPU.

wondering if this app is blind usable.

how to make LMStudio read the answer?

❤ Can it be used to make AI agents to run a small etsy shop? ❤❤❤

are you able to create images and surf the web?

Pro tip: max out the GPU slider on the right. To maximize speed you want to be able to fit the entire model into your VRAM, rule of thumb, the model should be 2GB smaller than the GPUs VRAM. The quants can be viewed as a compression technique, the smaller the number the lower the quality of the model. Generally, 4Q is a nice sweet spot for testing models and 5Q almost gives full quality outputs. This isn't always the case but following this rules you'll have a good time. Bonus point if you can make the standard Llama 3 take you to the darkweb. Have fun.

My use case: House Mind. I basically want my own Jarvis. Multiple personalities so I can switch from Spanish tutor to Math tutor to computer file sorter/duplicate finder, and always have control of house lights, appliances etc. I'll wait a while for the holography suite😜 #FeedTheAlgorithm

Set "GPU offload" to max. You'll see a BIG increase of speed ;) I think you have a 3090 or 4090 right? I always put it on max with my 3090, and it generates so much faster

@temp911Luke

Ай бұрын

Howdy, how many tokens/sec do you get when using Q4 or Q5 model ?

Great stuff many thanks

Seems that you like to play "The World ends with you" on DS ;-)

Ask it technical questions about Comfyui, does it answer them ok ?!

It hasn't build in with search Docs and search engines.

Can this be used on a laptop?

@InnocentiusLacrimosa

Ай бұрын

Yes. If it has good hardware.

i used llm studio a lot, but i cant load a visual model as yet succesfully. this would be hugh. not generating images more recockniton. If you have it working. Would be a nice video.

@okolenmi7511

Ай бұрын

You can run Stable Diffusion and some other types of models in ComfyUI. If you want to run only stable diffusion models you can use Automatic1111's UI - it have more "user friendly" web interface, but also it's not so optimized as ComfyUI.

Ай бұрын

@@okolenmi7511 yes i do this as well. What i didnt't could ran was Image recogmition Like vllava

Still wanting for more A.i based games. T.T

You forgot to set "GPU Offload" to MAX, hence you get barely 9 tokens/sec. On 4060 you will get between 30-42 tokens/sec (Q4/Q5)

The license says they use your CPU power when not in use for CRYPTO.... check what you're signing up for

@ASENDOMUSIC

Ай бұрын

woah what?

For whatever reason LM Studio doesn't fully offload the models to your GPU by default. You have to increase the layers to offload to max yourself. It can be so much faster!!

I would like to integrate it if possible to my game Cursed Dungeon Raider so that you could chat with the NPCs at the Black Market or the Historical Museum. But probably not important quest relevant ones. Moreover, an extra 5 GB RAM while running the game itself can be too much for some people's PC. Nevertheless, I really like to test that out. Let's see if this is possible with that.

I tried to use these LLMs but they are light years behind Claude or ChatGPT.

@bluesailormercury

Ай бұрын

Bigger models like Mixtral 8x7B or Llama 3 70B require more resources (RAM or VRAM) but are not far from GPT 3.5. 8B models are indeed too small for that

@InnocentiusLacrimosa

Ай бұрын

​@@bluesailormercuryindeed. 8B models are pretty disappointing for larger use cases. I need better hardware or more patience 😁

Greetings

It just needs a pdf reader :D

2:22 you mean GPU VRAM not ordinary RAM right?

@MattVidPro

Ай бұрын

It can be put on both but vram is much faster

@okolenmi7511

Ай бұрын

It's RAM. VRAM is another requirement. It depends on what are you using (GPU, CPU, Apple M chip, etc.). GPU is the most common case as it's fast enough, default CPU is the slowest option, but I'm not sure they implemented model work on CPU as this is not a good way to run LLMs.

Yo bro, suno 3.5 just dropped!

Zeh-'Fer if your are american, Zeh-fuh if you are a brit

@Earl_E_Burd

Ай бұрын

Yup, rhymes with heffer

Mlc-ai is great

Funny !

If you knew how to code man imagine the possibilities brother

it's VRAM not RAM right?

@okolenmi7511

Ай бұрын

RAM. VRAM settings in another place. You can set more VRAM to speed up your model. For example in this video were used 10GB VRAM to get that speed.

@UFOgamers

Ай бұрын

@@okolenmi7511 Thanks!

gpt4all just crashes with llama, i'll give this a shot

User: "...Oh my! Master Chief!" AI: "It's me, Patrick." Hmmm, not very impressive

@MattVidPro

Ай бұрын

better prompt would get the correct results - also for this example a completion tuned model would work much better (Not fine tuned for chat)

You have a powerful graphics card in your computer. When starting the model you made probably 1 mistake-you did not set GPU Offload to the maximum value. If someone has a weak graphics card and the model does not want to load-must turn this option off.

Explain to me, what do these llms really offer, when I could just ctrl-f the dataset itself or even make a simple software filtering program for the dataset, and likely get less hallucinations and other benefits. What is really being offered with an llm over the dataset itself.

@okolenmi7511

Ай бұрын

Good luck to make a program that will filter several trillions of words to get a single short answer. LLMs have some sort of creativity - not so much but it's enough to generate something that doesn't exist in dataset.

@apache937

Ай бұрын

go try that then

How about for iPhone yall ?

@joelface

Ай бұрын

Honestly I think the next iPhone is going to be custom built to run a local model about on par with llama-3 that will be able to control all of your apps and understand all of your requests with ease.

fuck yeah it works blindly easily for whatever hehe

Is it private though? Is our data safe?

Tbh, Ollama is better, because it's fastest LLM backend out of the box.

clone...still gonna try

Content crunch eh?