Deepseek 4.1 Flash on 2 DGX Sparks changed my life

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Deepseek 4.1 Flash on 2 DGX Sparks changed my life · Channel: Alex Finn · 16:36 · Watch on YouTube

Topics: AI agents, Software engineering with AI, AI infrastructure and hardware, Local and open-source models

Summary

Alex Finn buys a second NVIDIA DGX Spark, connects it to his first unit, and uses an AI agent to set up the two-machine system and run a local DeepSeek model. He demonstrates using the model through Hermes, including its response speed and its ability to create a simple Snake game and a more elaborate 3D roguelike. He argues that local models can be useful even when they are not frontier-level, because they can run continuously without per-use fees. The video encourages viewers to try local AI on hardware they already own, while also thanking Micro Center for sponsoring the video.

Key points

The video tests what AI models can run when two DGX Sparks are linked together.
The presenter says each Spark has 128 GB of unified memory and describes linking them to share memory for larger models.
He uses Tailscale to let devices in his home AI lab access the Spark cluster remotely.
He relies on an AI agent, Grokbot, to configure the devices and download a 400 GB DeepSeek model.
He connects the local model to Hermes and demonstrates sending it prompts through a new profile.
The model answers a short greeting and builds a Snake game in under a minute, according to the presenter.
A more complex 3D roguelike game takes about 20–25 minutes to create and test.
The presenter describes always-on, unlimited local-model use as “ambient intelligence,” with examples including web research and code review.
He says viewers can experiment with local models on less powerful hardware, such as a Mac mini, and ask an AI agent which model to run.

Chapters

Start	Chapter	Summary
00:00	Picking up a second DGX Spark	Alex introduces the plan to link two DGX Sparks and test the AI models they can run.
01:01	Unboxing and connecting the Sparks	He unboxes the second unit, explains its setup, and connects it to the first Spark.
03:37	Setting up the home AI network	Using Grokbot and Tailscale, he configures the Sparks so other devices can access the cluster.
06:43	Loading and testing DeepSeek	He asks an AI agent to install DeepSeek, connects it to Hermes, and tests its replies and coding ability.
12:56	Building a 3D game locally	He gives the model a detailed prompt for a Diablo-like roguelike and tries the resulting game.
15:00	Why use local models	Alex explains how free, unlimited local inference can support ongoing tasks and recommends trying local AI.

Quotes

We’re going to find out. — 00:00

This is a much better way of using AI. — 05:40

We are less than a minute in the code is written. — 12:25

It’s what I call ambient intelligence. — 15:32

When you have unlimited tokens, what you can do with those tokens change. — 16:03

Concepts and entities

DGX Spark, DeepSeek V4.1 Flash, Grokbot, Tailscale, Hermes, NVIDIA, Micro Center, Alex Finn, modelos locais, ambient intelligence

Full transcript (original language)

00:00 Today we’re at Micro Center to pick up my second DGX Spark. I already got one. We’re going to pick up a second, bring it home, I’m going to link them together, and we’re going to see what kind of AI you can run on two DGX Sparks. Can you cancel your cloud subscriptions by having two DGX Sparks? How powerful is the intelligence? Can you run your entire coding workflow? We’re going to find out. Thank you to Micro Center for sponsoring the video. They’re not my only favorite store, but they are the best electronic store on planet Earth. If you haven’t checked them out yet, go there, go online if you have to. They’re incredible. They’re the

00:30 smartest computer people there is. Now, let’s get in there and pick up our second DGX Spark. Oh, there it is right there. So, we are going to pick this up. We’re going to go home. I’m going to link two of these together. For those who don’t know, DGX Spark is like the easiest way to run local AI models humanly possible. If you link them together, you can have bigger and bigger clusters run more intelligent models. So, let’s pick this up, go home, I’ll show you how to link them up, and all the AI intelligence you can run on these machines. Let’s go. All right, we are back. Here it is. We are going to get this unboxed. We are going to add it

01:01 to my home AI lab, which consists of, let’s see, a Mac Studio, AMD Halo, the other DGX Spark, which we’ll be connecting it to, a couple of Mac Minis. We go over here, boom, two Mac Studios that will be on the network as well. By the way, this is This is what the studio looks like. Bit of a mess, please ignore it. All right, so let’s get this unboxed. Then we are going to connect them together, those two DGX Sparks. Then we’re going to load a bunch of AI models onto. I think we’re going to go with the new Deep Seek Flash, which I hear is incredible on two DGX Sparks.

01:32 Let’s get this unboxed. All right, as you see, we have two pull tabs. Let’s get those off there. Let’s see what we got here. Let’s get this sucker out of here. All right, boom. There it is, the DGX Spark. I like the presentation, very Apple-esque. Let’s get this out. For those who haven’t seen it yet, it is a beautiful device designed by Nvidia here. Very small, very easy to carry, very light, and it packs a ton of power. 128 GB of unified memory, an absolutely

02:06 incredible powerhouse of an AI computer. Doesn’t come with HDMI or anything. That’s because you don’t technically need to connect it uh to a monitor. You don’t need to do that at all. If you actually just plug the Spark into the wall and turn it on, it creates its own Wi-Fi network so that you can connect to it from a different device and set it up through like your main computer, which is what we’re going to do with my Mac Studios in just a second. We are going to connect uh to the Wi-Fi network that

02:36 the Spark makes so we can just set up through our Mac Studio instead of having to connect it to the monitor. So, let’s do this. I’m going to get this plugged into the wall. Then, I’m going to use this cable here. What this cable is is actually a connector that you can use that goes in the back there over on the left-hand side. You can see on the Spark. Connects into there and will connect to the back side of my first original DGX Spark that I have so that I can pull the memory together. 128 here, 128 here. That is 256 if my math is

03:07 correct that we can share between these two devices so that we can load some pretty big AI models onto them. So, let me do this. I’m going to plug it into the wall. Then, we’re going to go back to my original camera so that we can go on the Mac Studio and set it up. Then, we will load some awesome AI models onto it and we will do some really incredible things with local AI. Your mind’s about to be blown. All right. So, we got the second Spark all set up. We plugged into all, plugged into the other Spark. Set up was super easy. All you had to do was

03:37 connect it to a Wi-Fi network that was on this little piece of paper that comes with your Spark. You connect it to the Wi-Fi network. It runs all the setup. No need to plug it into a monitor or anything. From here, we can do all the advanced setup and configuration ourselves. I’m going to use Grokbot. You can use whatever AI agent you want. And what I did was is we made sure that everything was connected to the same Tailscale network. For those who don’t have Tailscale set up yet, if you have more than one device you own, you need to be using Tailscale. It’s completely free. This is not sponsored. It allows

04:09 you to create your own kind of private network between all your devices. Allows them to communicate with each other very easily, access them from anywhere in the world. You need to set that up. We’re going to get the second Spark onto that Tailscale network, so they can all communicate with each other. All right, so let’s set this up with Grok Bot. By the way, uh when you set up your Spark, it gives you this awesome like recipe book with all these different things you can do automatically uh with your DGX Spark. It makes it It’s like totally beginner-friendly, by the way. If you

04:39 just care about AI at all, you could buy one of these DGX Sparks and like do anything you want. Train models, train LoRAs, do whatever you want. It’s super easy uh with the the different recipe books they give you here. I generate images locally, all that. Super super easy. But, we’re going to do everything with Grok Bot. Just much easier that way, in my opinion. So, I’m going to give it access to my second DGX Spark. I’m going to give it my credentials, and we’ll get this all set up. All right, so Grok Bot successfully set up everything for me, set up the Tailscale network,

05:10 got the second DGX Spark onto the Tailscale network, so that now, not only can Grok Bot, but all my other computers in my home AI lab, my Mac Studios, my Mac Minis, the AMD, my gaming computer, can all contact that DGX Spark cluster, which is sick. This is great, right? This is the way you want to be setting it up, because now you can run those AI models on your DGX Sparks, whether you have one, whether you have multiple, whether you have a Mac Studio, whatever you’re running your AI on, you can have

05:40 it running on those devices, and then when you’re on the go, whether you’re on your phone, your iPad, your laptop, whatever it is, you can connect to those devices and use those local models. This is a much better way of using AI. It’s kind of having a home setup that’s always on, that never turns off, that’s on good computers, and then using nodes, like your iPhone or laptops or whatever, to use the models from those devices on the go. They connect to the Tailscale network, then they’re able to send

06:11 prompts to those home computers and get responses back. I think that’s a lot better than like taking laptops, like buying like really expensive laptops. You’d be buying like $10,000 like MacBook Pros with like 128 gigs of RAM. I don’t get it because you’re paying like a premium for like the laptop form and it turns off like every time you close your laptop and it gets super hot. Why not just have a home computer that’s always on that is like much better hardware for the price and you can just connect to it on like your cheap laptop

06:43 when you’re on the go. I think that’s a much better way of doing it. Anyway, a little bit of a tangent there. Let’s start getting this set up. Let’s first load models onto this. The easiest way to load models onto these devices now that we got them hooked up to each other is just asking our Grokbot to do it. So, if you don’t know which models to run on whatever devices you have. This isn’t just for the DGX Sparks. We’re about to put these two DGX Sparks to the test to see what kind of models they run, but even if you have like a Mac mini or like a cheaper Mac Studio or any device at

07:14 all, one DGX Spark, it doesn’t matter. The best way to figure out what runs on your device is just to go to any of your AI agents, Hermes, Grokbot, Dodds, whatever and just say, “Hey, take a look at that computer, see what kind of hardware it has and let me know the best model we can load onto it.” I have an idea of what model I want to run because I saw this tweet from Ray Fernandez showing this crazy Deepseek V 4.1 Flash just having just like absolutely insane speeds on two DGX Sparks. So, I want to load this up. I want to do this and see

07:45 how good this performs on the DGX Spark and see if we can get this type of performance. So, all I’m going to do is give this tweet to my Grokbot and say, “Hey, load this up on my Sparks.” That’s how easy this is. You don’t need to be like an AI model genius to get this working. You just have to have AI agents. AI agents make everything super easy. As long as you have AI agents, you can do anything you want. There’s no gatekeeping. There’s no technical knowledge necessary for any of this. You just go to your agent and say, “Hey, this is what I want.” Or even if you don’t know what you want, say, “Hey, I don’t know what I want. Help me out here.” And it’ll just figure it out for

08:17 you. So, let’s get this loaded in here. So, I am going to have Rockbot set up this deep seek model that is super fast and runs perfectly on two DGX Sparks. It’s giving me some setup to do to set the GPU speeds at the right speed. So, it’s giving me commands. It’s making this super easy for me. There really is no complexity here. All right, the model is now downloading 400 GB. It is going to go on there. It’s set up the link cable for me Rockbot. It’s downloading the model directly on. Never needed to plug this into my monitor. It’s just doing it all for me. And very

08:48 shortly, it’ll be running and we’re going to test this out to see how good two DGX Sparks linked together are and what kind of AI models can you run. Is it going to be good enough where you can cancel your Claude and OpenAI subscriptions? We’re about to find out. Before we do that, leave a like down below if you enjoyed this. This is kind of brand new kind of content for me testing out different devices, seeing what AI models run on it. If this is helpful, if this is more of what you want to see, leave a like and let me know down below. Depending on the feedback, I’ll just keep buying more and more computers and showing you what kind

09:19 of AI models run on them. Also, I got the number one AI newsletter on planet Earth. It’s completely free. Link for that down below as well. Let’s see how this goes. All right, it’s all set. It’s all loaded up. Deep seek is up and running across both Sparks. I’m telling you, there’s no rocket science involved with this. He plugged things in, you go to your agent and you say, “Hey, this is what I want to do. Do it for me.” And it does it. There’s no more gatekeeping involved with tech stuff. You can do anything you want. You just go to your AI agent, say you want it done, and it

09:49 gets done. So, we got it up and running. It has its own API coming from the Sparks to my computer. So, now anything on my main Mac Studio here can connect to it. I went ahead and I already followed up and I said, “Hey, can you make it so my Hermes can access it from my Mac Studio?” Create a profile on my Hermes to talk to it. I highly recommend going with Hermes for this. If you want to build an agent on top of this local model you’re running, Hermes is free, open source, super easy to use, and built for use cases like this. Literally

10:21 less than a minute later, it was all set up. I have a new profile on my Hermes agent, good to go. So, here we go. Look at that, Deep Seek. I have a new bot, which is a new profile in Hermes, all set up where I can talk directly to my new local model. Anytime I want to talk to my local model, I can go click on Deep Seek and start talking to it. It’s completely free, it’s unlimited, it will never cost me a penny outside of the electricity going into those DGX sparks. So, let’s send in a couple test messages and see how it goes. Hey, are you there?

10:52 Let’s test this speed. Let’s see how good the speed is here. All right. So, we sent Hey, are you there? This is all real time. I’m not going to edit this out. You can even see the time counter under it. There we go. Yes, I’m here. What do you need? 8 seconds. Say hello to my YouTube audience for me. I’m going to hit enter on that. Let’s see how quick it responds. You can even see the thinking. That’s another great part about the local models is you can see the full thinking traces. A lot of the frontier models now obfuscate that, so you can’t see thinking. So, you can see

11:23 everything that happens inside your model, which is a really beautiful part about using local models is everything’s open. There’s nothing hidden, there’s nothing obfuscated. You can see how it thinks, how it works, all of that. Here’s a short greeting to drop into the video intro. Can you build me a quick snake game in a single HTML file? I’m going to hit enter on that. Let’s see how this is. We’re going to give it a much bigger, more intense test in a second to see how far we can push this. So, let’s just see how fast it can build

11:54 something really simple for us, and then we’ll get into the more intense build us some big 3D experience. So, it’s thinking. It’s been thinking for 20 seconds now. It’s already come up with the controls. Let’s see what else it’s thinking about here. It’s already writing the files. You see it’s already got the game logic going on, the implementation. We’re 20 seconds in, and it basically has the entire thing planned out. The code details. This is rapid rapid fire. I mean, this is so quick that I could 100% see myself using this as like a main model to do things.

12:25 Right, this was 6 months ago, 12 months ago, this would be frontier level running lightning quick all on my desk unlimited for free, which is mind-blowing. And boom, look at this. We are less than a minute in the code is written. Now, we’re going to see how well we can test this. All right, boom, it’s done. It’s nice. We got this nice artifact inside our Hermes chat here. So, we can actually play it inside here. Let’s go. Let’s test it out. All right, here we go. Oh, I mean, I’m going to be honest, for a snake game, it don’t look half bad. We got the corner. Okay, okay, okay, can I get the corner? Oh, no, I can’t. Can’t get the corner. Pretty

12:56 good. Built in like a minute and a half on a local model. Now, let’s push it to its limits. Let’s see how good we can get here. All right, so we’re going to have it build a pretty intense 3D experience here. I want to build a roguelike game using 3JS. I want to be a top-down demon slaying game like Diablo where each level you get more and more demons, but in between levels you can choose a specific power to get or attribute to level up. It should be a dark demon-like similar to Diablo in fun satisfying gameplay with in-depth detailed beautiful stylistic graphics.

13:26 That’s pretty intense. We’re going to have it do it though. I’m going to hit send on this and we’re going to see how it goes. I’m going to put this prompt down below as well so you can just copy it if you want. I’m super into building video games right now. I think it’s absolute blast to build. So, feel free to take that build your own version. Edit any way you want, tweak it any way you want, you’ll be good to go. Fun exercise, do with me. Can’t hurt. All right, we are done here. To be honest, it took about 20-25 minutes. This is a pretty intense game. This is a big multi-level complex 3D game. Looks

13:56 like we have a bunch of screenshots it sent us, which is nice. It tested itself, which is super nice. I mean, this is a local model testing itself, being smart, seeing how good the game is, sending us screenshots. All done locally, all done for free. Let’s test this out ourselves though. All right, so I had this open up in my browser. This is pretty sweet. The 3D graphics for a local model. Again, local, non-frontier, completely free, and unlimited model is pretty sick. The gameplay does feel like

14:27 Diablo. I mean, I don’t have like a ton of like spells or anything or classes like we have in Diablo, but let’s see how far we Okay, now we can add attributes. Reflect 25% of damage taken, heal for 12% of damage dealt. I feel like you got to get that vampirism. Okay, now we’re back in it. My guy’s looking I think they added onto his armor now. Let’s see how we do here. Let’s use one of our abilities. Boom! Okay, we got bigger bosses, bigger enemies. That is sick. I mean, this is really for one pass, one shot, one prompt game, right? Like, this

15:00 is pretty impressive. There’s crits. You can see the large numbers going up. Like, now I can go back and keep working on this and improving it with the model. Again, completely for free. I think a question I get a ton is like, “Okay, Alex, that’s great, but it’s not frontier. So, when would I ever use that? I pay $200 a month for chat GPT. Why would I use this model?” Well, the use cases change when you have an unlimited, for free, model, right? Yes, if you’re doing like super tough coding, you’ll use your subscriptions you got

15:32 with Claude, Opus, whatever. But, because this is free and unlimited, there’s a lot of different things you can do here. It’s what I call ambient intelligence. Things you can just have going around the clock. For instance, I have my local model going and like searching the web for business opportunities all times. Looking for ways I can build businesses and make money. I have it looking over my code bases, reviewing code over and over and over again throughout the day. These are things if you did with cloud models, all your usage would be eaten really quick. But, because they’re local, I can just

16:03 have it going. When you have unlimited tokens, what you can do with those tokens change. So, I highly, highly, highly recommend getting started with local models. You don’t need two DGX Sparks to do this. Although, having two of them is awesome and allows you to really, really sick things like this. But even if you just have a Mac mini, man, you can just go go to a Hermes agent say, “Hey, what’s the best local model I can run in this Mac mini?” And it will download it and run it for you. It really is that simple and it’s something everybody should be doing. Is it the best models in the world? No, but

16:36 it’s your own. You learn, it’s education, and you get better at using it and you can unlock some really sick use cases that go around the clock. Highly recommend doing it. Again, thanks to Micro Center for sponsoring the video. What great partners. Support local retail. Go to Micro Center. Hope this was helpful. I’ll see you in the next video.