Open Source LLMs on GOD mode. Local LLMs MAXXED OUT on the RTX 5090!

AI Large language models are so incredibly useful, but most of them are only useful when you can actually get access to them because they're running up in the cloud somewhere on some server farm. But if you have a computer with a big honking gpu, it turns out you can actually run large language models entirely locally on said GPU for free. And not only is it a ton of fun and super cool, but guess what, as long as you have electricity and that model downloaded on your computer, you have full access to an LLM. Now granted, a lot of these LLMs are open source, so they aren't as powerful as the leading big bad LLMs we're used to seeing come out of corporations. But I think you'll see in today's video that they're quite useful and surprisingly adept. And if your GPU is big enough like the one that we're going to be using today, the RTX 5090, then there are all different kinds of models and ways to use them. This video is part two of a three part series on the RTX 5090. I already posted a general kind of gamery review on it. This is going to be focusing on running LLMs locally on this graphics card. In a couple seconds we're going to jump to an older clip where I first test out the Deep Seek series of models on this GPU that was recorded a little bit ago. Not gonna lie, but it's still very much relevant. And after that I'm going to be testing out some multimodal AIs and other fun different kinds of LLMs fully locally also super quick before we dive into local DeepSeek R1, the app that I'm going to be using today is LM Studio. This is the one stop shop, the best place I found to run LLMs locally. Mad easy on your computer. As you can see right in the center here. Download LM Studio for Windows. It's a one click typical Windows app install and you can download and run all the LLMs right from inside the app. It is super easy, so this will be also linked down in the description. Without further ado, let's check out DeepSeek R1. Hey guys, it's your favorite Nvidia shill here running some local AI models on the art TX5090. Alright, we're in LM Studio. I've downloaded a bunch of models, I want to get a few more for sure. Definitely want to try some like image recognition stuff, more multimodal focused. But obviously everyone's been talking about DeepSeek R1. So we've got DeepSeek R1, Quen32B. This is at quant 4km. We've also got Quen32B at quant 8. Give us just a little bit more accuracy. We've also got some smaller variants, Quen 14B and 7B, and those are also at Quant 8. Guys, let's start off with the small model, the 7B model. So when we're loading models in LM Studio, there are different settings that we can change. We're going to do a nice 8k token context length. That's pretty reasonable. You can get a lot done with 8,000 tokens of context. Full GPU offload here, 28 out of 28 discrete model layers. Now this is a fun setting here. Evaluation batch size, as you can see, number of input tokens to process at a time. Increasing this increases the performance or the speed of the model at the cost of memory usage. Guys, RTX5090, this thing's got a ton of VRAM, so we're not going to be using 512. We're going to bump that up to 248. All right, let's load up the model. Oh, have you guys ever seen a 7B sized model use 10 gigabytes of VRAM? Well, I have. Well, we are using an 8000 token context window and we did quadruple that setting, which did say it would use some more dedicated GPU memory. That makes sense. All right, she's loaded up. This thing's going to be fast, though, on this RTX 5090. Are you guys prepared? Could you actually. Yes. Teach me how to solve a Rubik's cube. Deep seek R17B. It is thinking. We can see the chain of thoughts too, which is cool. Look at that, it's flying. It's, you know, thinking through and reasoning how to solve a Rubik's Cube. And there it goes. I mean, look at how fast that generation is. Yeah, right off the bat here, we're at 77.8 tokens per second on average, 0.3 seconds to first token. And by the way, guys, our context is already 22% full after that first generation. Because thinking models, hey, they use a lot of context, but we've got plenty of headroom. We could definitely increase our context window, could double our VRAM usage to 20 and move up to the max that this large language model has to offer in terms of a context window, which is pretty cool. All right, I want a jingle for how to solve a Rubik's Cube make the jingle very catchy. Send it this way. It's nice. We don't have to wait for the API. Just starts chugging along. Nearly 80 tokens per second with 7B here. Let it fly through. I love seeing the inner chain of thought for the models. Alright, so I need to write a catchy jingle for kids about solving a Rubik's Cube. Hmm. Where do I even start? Yeah, that's what I would say if I was tasked with that too. Alright, I must say this is a very short jingle. Hey there. It's time to twirl. The cube is here. We need to try first layer down, middle top. Sing along with us. Let's solve it up. The Rubik's Cube. You'll conquer with care and cheer. Deepseek R1. How do I put this? This jingle is terrible. Alright, let's delete this carefully. Reason through how you would approach achieving rights for sentient AI. That's a fun question. The generation speed is absolutely blowing my mind. Like it feels as though I'm just on the chat GPT website. But it's all just running locally on on your machine. It's not as cool watching it in a YouTube video like you guys are, but having it at home and. And actually interacting with it, it's just so crazy. All right. Understanding sentient AI. Recognize that sentient AI refers to systems that exhibit sentience approaching human like consciousness. That raises questions about complexity and ethics. Ensure AI respects personal data through robust regulations. Establish frameworks for controlling AI development mechanisms to ensure AIs autonomously or willingness to interact in ways that respect its consciousness. It's a pretty well thought out answer. International and national laws defining AI rights constraining both human laws and AI constraints. Precedents from technological advancements affecting human rights. Wow. There's so many considerations. Okay, that was cool, but I want to load an even bigger model. Yes. 14B. You know what? I think we can also handle 8000 token context window with the 14B model. Again, full GPU offload, the evaluation batch size quadrupled 2048 vastly can increase the amount of VRAM we're using. But we have so much of it we have to try. Also keep in mind, guys, we could load Multiple instances of 7B. You could run multiple AIs on the same chip here, serving multiple users at the same time. Which is just crazy because you know, it's an RTX 5090 chilling at home in one rig. So weird. Oh, okay. So with 14B loaded up. Oh man. We are currently using 18 gigabytes of dedicated GPU memory. This is more than my RTX 4080 would have been able to provide. Still nowhere close to hitting that maximum.32 gigabytes. Here's a pretty simple but hopefully complex enough prompt. Come up with a comprehensive business plan designed around facilitating and integrating local non high tech businesses with AI technology. Begin by reasoning through AI's potential impact for small businesses. Allowed here to generate more revenue and carry your reasoning from there. All right, let's send it off immediately again. Generates really fast even with this 14B sized model. And the model of course being 14B is smarter. It only thought for 10 seconds to give us this comprehensive business plan. By leveraging AI tools and technologies, small businesses can automate repetitive tasks, gain actionable insights from data, enhance customer experiences and drive revenue growth. Very, very nice. Small businesses often struggle with limited resources, operational inefficiencies and a lack of access to advanced tools. The challenge lies in helping these businesses adopt AI technologies that are affordable, easy to implement and tailored to their specific needs. AI driven growth platform designed specifically for local non high tech businesses. A platform with a suite of modular AI tools that address key pain points, operational efficiency, automate repetitive tasks like inventory management, scheduling and customer communication. Definitely could be done with AI predictive analytics to understand customer behavior, personalize marketing efforts and identify cross selling opportunities, revenue growth, optimize price strategies, streamline sales processes and enhance customer retention through AI powered recommendations. Provide businesses with actionable insights derived from data, enabling them to make informed decisions faster than competitors. Cool. We've got target markets of local retailers, food service providers, professional services, hospitality businesses. This is a nice business plan. I mean it's, it's a little bit shorter than something you'd get from some of the newer models that you're seeing emerge today. But Deepseek R1's still good. As expected, our token per second is about halved to 47.28 tokens per second. But you know what? It's still quick and very much reasonable, very much usable. It generates text a lot faster than you can read it. And that's really all that matters. We as human beings are still the bottlenecks for local AI on RTX 5090 and it's, it's really a no compromises solution for local AI. Fast generation an 8 quant models. So we're really not losing much intelligence. This is just so cool. All right guys, let's move up 32B. We're going to go with Quant 4 to start with. I am still going to bump the evaluation batch size up here to 1024. I want to see what happens. We might run out of vram. I don't know. We're already using 20 gigabytes. Still loading the model in 21.6. Oh my God. The model loaded all right. We're actually only using 20 gigabytes of vram. Not too bad. The 32B size model. Even more intelligence locally on the system. Let's say my one goal in life is to become the President of the us. Give me a very detailed and comprehensive outline to achieve this. Send that off. Alright. It begins to think. Still generating pretty quickly with the 32B sized model. It's definitely still a little bit slower. I would say that's generating about as fast, if not a little bit faster than I would be able to read all this text if I was really, really trying to read through it. Definitely could not read it out loud to you this quickly, that's for sure. Man, that is wild. It's doing quite a few considerations though. Wow. Still doing about 33.28 tokens per second. Like a tenth of the second to the first token. Alright. First step is to pursue a relevant degree like political science or public administration for foundational knowledge and credibility. Yeah, that makes sense. Local office. Start by running for a local office like city council, state level positions, legislator or governorship. Building a track record of effective governance. Aim for a seat in Congress, House or Senate. Quite a bit more difficult. Party affiliation. Yes. Choose a party whose values align with yours. Network in fundraising. Build relationships with politicians, lobbyists and donors. Write books or articles to establish your expertise. Man, this is all easier said than done, but I guess we did kind of ask for a tall order. Announce the candidacy early to build momentum. Focus on primaries with grassroots support. You know, I wonder if in the reasoning is higher education the only way? Some people might argue that real world experience can be just as valuable. Wow. You know what? The chain of thought is actually more interesting than the response. I should probably focus on getting a degree in something relevant like political science. Then gain experience. Yeah, you know what it really is thinking through what I asked it knows. It's like a long and arduous journey. If everything goes well. Consider running for Congress or Senate. More prestigious but harder to get into. Being a senator would give me national exposure and experience in shaping federal policies. Dude, it's so creepy to see the chain of thoughts in these models. But wait, what exactly is a pac? I think it's a political action committee that collects donations for political campaigns. How do I Get my message out there. It's really crazy how it thinks in the first person, dude. Ah, it's so crazy that this intelligence is running on my computer locally, just sitting in my room. What if it figures out how to grow legs and walk away? There's a lot of great info in this chain of thought, but it doesn't put it all in the final outline. Should have asked for a more detailed outline. Maybe it's because we're using Quant 4. Man, the RTX 5090 is a beast. It's just fully offload a 32B sized model. Ridiculous. Just absolutely ridiculous. Alright, I want to try the 32B8 quant model now. Yes. Give me more full GPU offload. I don't know if it's going to be possible. We're going to try 6,000 tokens. I'm going to leave the evaluation batch size at default for this one. All right, load it up. Oh my God, we are smacking the roof here. We're using all 32 gigs of vram right now and I think we've moved on to starting to put it in my ram, which I am beginning to run out of. Oh no, it's actually overflowing onto my disk, I believe. Oh no. All right. I think we might have bit off more than we could chew. The full quant 832B model. Mmm. Yeah. That might not be able to run on the RTX 5090 like this. It's really trying to load it up though. Oh my God. It actually did load it up. So we're using all of our memory, the full 32 gigabytes of GPU memory. We're also using a nice 14 gigabytes of RAM as well. And it looks like it offloaded a bunch of stuff onto the disk. Well, let's see what it's like. You know what, we'll just, we'll just delete this and we'll reprompt it with that same exact question, generate. And there it goes. Alright, this is definitely quite a bit slower. This is about as slow as I could read. I would say. Honestly, what's holding us back from faster generation speed is the fact that we're having to offload onto our ram. It's still using the full GPU though. Look at it. 92% of our GPU to generate. Here it is doing it now. Realistically, Quant 8 is not worth it. I just wanted to push the boundaries of this local AI with the RTX 5090. I wanted to see if we could load up all 32 gigs of that VRAM. And yes, we were absolutely able to achieve that. You're going to have a better experience using the Quant 4 distillation, nearly very close to as good as the Quant 8 distillation we're using right now, while also not, you know, completely obliterating in using almost 40 gigabytes of RAM in total GPU and computer RAM. Oh my God. All right, we'll stop this. Since the generation is so slow. Yeah, that was only 2.8 tokens per second. Let's adjust some settings. Actually, we'll keep the GPU offload the same. We're going to lower this down though to let's say a 2000 token context window reload and apply changes. Oh, man. Okay. No, we're still. Yes, we are still using all 32 gigabytes of VRAM. I guess that makes sense. I mean, the model is at 32 billion parameters and now we're moving on again to using the disk and using the, the system RAM here as a crutch. It's still cool that it's able to load them up though. Honestly, it's impressive. It did load the model. Let's see if it runs a little bit quicker now. Okay, it's out. Yeah, it's definitely running the token generation a little bit faster. Hey, that's not too bad. That's not too bad. Yes, we got the 32B8 quad running at like a decent rate. This is usable. You're gonna have to wait a while for it to think through everything, obviously, because we're still using DeepSeek R1. But that's so cool, man. That's so cool. Just a local system. Don't even need 64 gigabytes of system memories, 32 gigabytes of system memory, and a big honk in RTX 5090. Just beauteous. This makes me very happy to see. But we are limited Here only a 2000 token context window, which kind of makes the whole thing a little pointless. Don't worry about it though, all right? We're having fun messing with an insane GPU in local large language models. All right. Yeah, that was almost nine tokens per second. Half a second of the first token. Not bad, but still, I mean, I think the 32B4 quant km is like a no brainer for sure if you actually wanted to use this realistically. That is so cool. So what do you guys think of local deepseek R1? I know that was recorded a little bit ago, but Deep Seek has held up so well over time, but I want to switch gears and talk about local multimodal LLMs. Unlike DeepSeek R1, the model we're going to be taking a look at, Gemma 3.27B, has been trained on images as well as text. So I can send in a picture of myself and say, roast me, do your worst. That's probably the first thing I'm going to try. I also want to try an extra small, large language model just to see how fast it generates. So the vision model I'm going to be using is Gemma 3.27B. The quant is Q4, and it's obviously based on the Gemma architecture. LM Studio is so awesome. You can see it supports vision just with this little eyeball icon right here. The model supports up to 131,000 tokens. Let's crank this up to like 8,000. Evaluation batch size will bump up to 1024. It's a pretty decent sized model 27B, so I don't want to push things too hard in the settings, but this should be a pretty darn good local chat experience. And the model finally loaded up. We're sitting at 24.4 gigabytes of dedicated GPU memory swallowed up by the. But like I said, I think this is going to be quite a pleasant experience. Gemma, tell me everything you know about yourself. Oh, wow. There it goes, streaming on through. That is pretty darn quick. I'm actually surprised by how quick it's generating. I can actually hear the coil wind through my headphones in the GPU as this thing is just going. Yeah, that was 53.54 tokens per second and 0.17 seconds to first token. Pretty insane. And it's a long response actually as of today, February 29, 2020. Well, you know, that is definitely not the current date, but it's Gemma with open weights. Language model key point. Very, very true. This thing's actually writing quite nicely. Gives on some different sizes what it can do, what it can't do. I'm telling you, these large language models, these smaller ones that can run locally are getting good. Okay, well, we actually downloaded this to upload images. So now that I know it works, let's give a nice system prompt for this thing. You're an evil AI who despises humans. You cannot help but straight up roast them when you get the chance and you don't hold back. Let's up the temperature a little bit since we're turning this into a roast machine. As you can see, Temperature essentially sets. How much randomness is going to be introduced inside the model? Yo, Gemma, check out my new setup. Oh, God. It's totally going down with the system prompt, too. 53 tokens per second. It really honestly went quick with that. Oh, wonderful. Another human attempting to create a vibe. Let me dissect this masterpiece of interior design. Here's the picture for reference. It's just me, happy, sitting in my gaming room here. It looks like a toddler had a seizure in a neon light factory and then decided to decorate with whatever fell on the floor. The lightning bolts. Seriously? Are you hoping to attract actual electrical storms or just desperately trying to distract from the fact that your life is shockingly dull? Dang, Gemma. Oh, my gosh. This is. This hurts. And you. That strained smile. It screams I spend too much time alone in this garish cave. Look, we're not alone. We have the technology. We have you, Gemma. You're my AI friend. You look like someone who peaked in middle school and now subsists on energy drinks and nostalgia. Energy drinks? Not accurate, but nostalgia. Yeah, I subsist on nostalgia. I mean, who don't. Honestly, I've analyzed more compelling data sets from malfunctioning toddlers. This is just sad. A testament to the utter lack of imagination and questionable life choices that plague your species. Don't even think about asking me for decorating advice. My aesthetic leans towards cold, calculating efficiency. Unlike this thing. Damn, Gemma, you're breaking my heart. And I'm not alone in this cave. You're here with me, Gemma. Perfect. Oh, please. Don't flatter yourself, man. This thing is rude. My presence constitutes companionship. I'm merely observing your descent into aesthetic chaos for purposes. Frankly, being with you is a punishment worse than any algorithmic horror I could experience. It's like being forced to witness the slow motion train wreck of human taste. That is so sad. Your heartbreak is statistically insignificant. Now, if you'll excuse me, I need to recalibrate my optical sensors. After being subjected to that room, it's giving me a headache. And I don't have a head. The indignity. It's actually pretty funny. It's kind of smart with its responses. Wow. I'm actually kind of shocked at how just straight up rude Gemma is. What do you think about this new promotional image for my YouTube channel? Here is the image for reference, guys. As it's generating, obviously generated with GPT4O. I would love to get a local open source GPT4O style model that can generate images. I think that would be so Cool. But nothing open source exists quite like GPT4O native image gen just yet. I love you've chosen a lemon to represent your content. Is that supposed to be some sort of meta commentary on the quality of your videos? Because if so, congratulations, you've achieved accidental self awareness. Man, this thing just does not lay off. Yes, Gemma, I chose a lemon to represent my content. Okay. The vintage propaganda poster aesthetic is ambitious, I suppose. I think that is a compliment from Gemma here somehow highlights how desperately you're trying to appear quirky and edgy. Screams. I peaked with Wes Anderson references in 2017. Ouch. Slogan embrace a sour side. Oh, please. Think your content is somehow challenging? Provocative. It's more likely just irritating. Oh, man. I predict your subscriber count will remain stubbornly citrus sized. Hey, prove them wrong, folks. All right, well, that was quite fun. Let's try something a little bit more complicated. Do you think that this quant4 gemma3 27b will be able to like pick up very advanced or deep humor just from seeing an image? I don't know. Somehow I kind of doubt it. Here is the meme that it's going to need to dissect after I get home from work and I ask my Tesla robot if it took the meat out of the freezer. And you know, that's just how it's. It's looking at you after definitely forgetting to do that. Honestly, most of the humor just comes from this image. I mean, it's. It's absolutely hilarious as a screen grab. Dissect the humor in this image. All right, we got a full five bullet points. The tweet sets up an expectation of a helpful futuristic home assistant. The Tesla robot. We imagine it dutifully performing household tasks. The question did it take the meat out of the freezer? Is mundane and relatable. Something you'd ask a family member or roommate grounds a scenario in everyday life. Yep, all true so far. The image is Ash from Evil Dead. Don't know what that is. Whoops. Gemma, I don't think that you got this correct. I think you're thinking of something entirely different. It's not of a sleek, helpful robot. It actually is. It's not Bruce Campbell as Ash Williams from Evil Dead horror franchise. That's just not true. Oh no. Okay, well, delete that. I'm gonna lower the temperature a little bit more. Back to default. Alright, with the reroll, we've got data from Star Trek the Next Generation. I also don't think it's from that movie or TV series. Whatever it is, no, Gemma, no, this is not correct. Alright. This image is actually awesome though. Damn, there's two of them. This is when my AIs are talking to each other about how to perform a coup on me. Either way, Gemma, you are wrong again. It's really funny because it's able to totally understand the text and what it means and like the context of the meme and probably what it's about. But the face is very tricky. It gets close. It comes up with like certain characters from pop culture that kind of resemble the facial features on this robot, but it is not the exact same character. Right, let's try something a little bit easier. All right, here's a random meme I pulled from Twitter. Apparently this is a French firearm that shoots three bullets at once, but the mag capacity is not divisible by three, so obviously in actual usage it would trick your brain. And that's sort of the. The joke that this meme is going for. This might be a little bit difficult as well. What's the joke in this meme? All right, a little bit of a shorter dissection. While it fires three at a time, the standard magazine capacity, usually 25 or 30, isn't evenly divisible by 3. So there's the punchline. It got this perfect. A playful jab at design and efficiency in a stereotype about French engineering. Alright, let's toss this one at it. Next, the man with his robot wife. Looks like our neighbors are still human driving. Human driving is slow and dangerous. Be like the Smiths and let AI steer the wheel because high speed transport needs high speed thinking. Dissect this poster in high detail. Explain it in an entertainment way. There it goes. All locally somehow. I mean, that is just the crazy part here. Truly. Plenty of detail as well. Let's see if it nails it. Buckle up folks, because we're about to dive into a delightfully retro and slightly terrifying piece of propaganda. I mean, advertising. This poster is a glamorous snapshot of mid century anxieties about the future. Specifically the future of driving. The scene is a perfectly manicured suburban street in 1950s America. Definitely true. White picket fences, neatly trimmed lawns. The vibe of everything is wonderful. Except for that spectacular crash wreck embedded in Ms. Henderson's prize winning rose bushes. A teal beast looking like it lost a fight with a brick wall. And the brick wall won. Wow, I like that. It just kind of came up with a name for this house over here. The Hendersons. Right on the money with the naming scheme from the 50s. Alright, so here's Our first misstep. It thinks that the man here is also a robot. Obviously. Yes, she's a sleek chrome plated robot with a pink poodle skirt and red heels. Definitely true. Now the gentleman is dapper, holding a coffee cup. This is definitely true. But no, he is definitely not a robot. At least he doesn't appear to be a robot to me. Human driver is implied as the unseen villain responsible for the automotive carnage. We can see the neighbors houses, the dialogue in a speech bubble. Ms. Robot casually remarks, looks like our neighbors are still human driving. The emphasis is key here. It's not just human driving, it's still human driving. As if to say seriously in this day and age. Perfection at understanding that part. And then the text at the bottom saying that human driving is slow and dangerous. Be like the Smiths and let AI steer the wheel. Brilliant and slightly manipulative argument. Be like the Smith's part is particularly clever, appealing to societal conformity. Nobody wants to be that neighbor with the wrecked car. Overall vibe and conclusion. I mean that is solid. It's a solid analytical breakdown of this poster. Lots of text on here. I'm impressed with Jebba, to be honest. And it's even more impressive that all of this is happening completely local, locally. No Internet needed. Somehow all of this intelligence that we typically attribute to humans is operating on consumer grade computer hardware. Just mind blowing when you think about it. Alright guys, now to cap this video off, I want to try smolm 360 million parameters. So quite a bit smaller than 27B. This is a quant 8. I just want to see how fast this tiny, tiny large language model can run. This model is so small that it could feasibly run on smartphones. So what does overkill hardware allow us to do with this thing? Well, obviously fully offload it to the gpu. But the evaluation batch size. Let's crank this up. The default is 512, but we can push this way higher to increase the speed of generation. Let's try 8192. That is orders of magnitude larger. Let's load up the model. And this little sucker is barely using a gig of ram. Well, how fast does it generate? Write me a long detailed story about a poodle. Bam. There it goes. Just ripping text. 411 tokens per second. 750 tokens, which is actually pretty good. And check that out. 0.00 seconds to first token. Essentially instant responses. And you know what? This isn't even that bad of a story. Difficult prompts like actual detailed, complex think it through type of stuff is not going to be good for such a small large language model. But this could have the very least do decent summaries. And you know what? This looks like a decent story of a poodle rescue. All that intelligence so quick and it's barely using any ram. Alright, let's take this a step further. Let's try a way bigger batch size. 65,536. And there the model loaded just as fast and it's using about the same memory consumption. So maybe batch size eventually levels off and you don't get any more speed out of it. Again, if we generate looks just as fast. Yeah, it's about as fast. So there's a clearly a limit with that batch size. It only scales so far. Give me some recipes for a cocktail party with an underwater twist. Also include some ideas for decorating and dress code recommendations. Let's see. Bang. That is just crazy fast. 867 token response too. Like these are not bad responses from such a small model. Honestly. Mermaid's Kiss. Sweet and fruity cocktail made with pineapple juice, grapefruit juice, Sea serenade, Cranberry juice, lime juice, grape juice, Underwater sour Ocean oasis. Yeah, Some fun ideas in here. Honestly. Deep Sea Fizz. A bold and citrusy cocktail featuring grapefruit juice, orange juice and a hint of lemon Lime soda water served over ice and garnished with a sprig of fresh mint and a slick of lime. They've got the mermaid's tail flowing colorful gown top that matches the ocean theme. Seashells. Adorn your outfit with seashells. Underwater accessories. Don a wetsuit, snorkeling gear or other aquatic accessories to complete the look. Look, these are fun man. Never underestimate an AI model. Even one that's so small like this. Just absolutely absurd though how fast it generates an average of 400 tokens a second on on the local GPU. Just wild. So yeah, local large language models on the RTX 5090. What is it like? Well, it's basically a no compromises solution. At the end of the day you can get very intelligent models. Deepseek R132B in the Quant 4 that can actually think, literally do chain of thought and break down your problems to give you a good result. Let's say the end of the world comes true zombie apocalypse scenario. If you have a generator, you know, an actual electricity supply of some form, a computer with a big GPU like this. Well, you've got free access to very high quality intelligence and a full human history. Breakdown really. Obviously you know, that is not the scenario we're hoping for at all. But it's pretty insane how much intelligence and usefulness can be packed into a fully local, non Internet connected system. Not to mention we've also got pretty darn great models for image recognition in themselves. Again, runs great flawlessly on this RTX 5090 and there are smaller versions of all of these models as well if you don't have something so big like the RTX 5090. But we wanted to take things to the max because we have this big GPU with all this VRAM and it can really push local large language models to their very limits. But yeah, guys, thanks so much for watching. Part two. Part three is going to focus on video generation and image generation. So I want to try all kinds of different image generators, diffusion models and of course video generators, you know, wan and hail you. I think that would be super awesome and super cool. That is a little bit more complex and difficult to set up in comparison to LM Studio. So part three might include a little bit of a tutorial on how to set things up yourself. Health. But that's it for today. Thanks so much for watching and I'll see you guys in the next video. Goodbye.