i got one....and it's FAST!!!

0:00 I got one. An M5 Ultra. You wanna see it? It's empty. I'm a terrible unboxer. Now I did that to illustrate the fact that I'm not gonna use this like you would. I'm not gonna put this on my desk and plug my keyboard mouse into it, my monitors. No. No. No. It's gonna be in my server room because I treat it like a server. Look. There it is. All I care about is local AI on this sucker, and that's what we're testing in this video.

0:27 It's supposed to be four times faster than its predecessor, the M3 Ultra. I also have one of those. Let's see how they do head to head. Yeah. We're gonna do benchmarks. Don't worry. But it's not just that. We're gonna really test it for things I might use it for. Image and video generation, transcription and video ingestion. What about Kev, the open source version of Jev or Laya? Can a local model use DaVinci Resolve to edit my videos?

0:50 We're gonna test all that. Get your coffee ready. Let's go. And shout out to Micro Center for sponsoring this video. We'll talk more about them later. And before we start, couple of ground rules. Apple loaned me the M5, not mine. They sent it to me. They said, look at this. And I said, okay. They don't control what I say, but they do control me sending it back, which makes me sad. Number two, we're running local AI on this thing.

1:12 You're not talking to OpenAI or Anthropic servers. Everything's happening right here. No Internet required. But these are all my tests and all my numbers. That's it. Let's go. Now here's what I want you to pay attention to. Here's the spec comparison CPU. Yeah. More cores. Cool. But there's two things I want you to focus on. First, the GPU. We had the same 80 cores in both the M3 Ultra and the M5 Ultra, but every one of these cores has a Neural Accelerator, the M5 Ultra.

1:42 What's it accelerating? We're about to find out. And then this right here, the memory bandwidth going from 819 gigabytes per second to 1.2 terabytes per second. So memory bandwidth and GPU core neural accelerators. Let's watch it happen. Okay. The race is set. Right here, we are rocking the Qwen3.8 27B 4-bit model on both machines. I'm gonna switch this sucker over to a 32K prompt, and here's what we're actually doing. I'm sending a 32,000 token essay to both machines and saying, hey.

2:13 Summarize this sucker. You ready? Let the race begin. Start. And they're off. Perfect time for a coffee break. I wonder who's gonna win. It's gonna be the M5. But by how much? Dude, he finished. Okay. Good job, M5. And a couple of cool things happened right here. I I I cannot wait to tell you what it is, but we're gonna wait for the M3 to catch up. Come on, M3. There he goes. Good job, M3. Clearly, we have a winner.

2:41 And let me show you my favorite metric. Look at this. This right here, the first word, M5 did 21.66 seconds, the M3 did 82.85 seconds. What does that mean? I like to call this the wait. The industry might call it TTFT or time to first token. But what does that mean? Well, when you prompt these models, when you send that 32,000 token prompt, it has to read the entire prompt all at once, all 32,000 tokens. This is hard to do.

3:07 That's some big brain matrix multiplication math and requires a ton of GPU power, but the M5 smoked the M3. Why? They both have 80 cores. Right? 80 GPU cores. What's the difference? It's probably gonna be those neural accelerators. Right? That's what I'm guessing. They accelerated that thing. Looking at our scoreboard, the M5 was 3.8 times faster. And if I scroll down just a bit more, we can see where else the M5 won. On speed, it was 1.5 times faster.

3:35 And when I say speed, it's up here. 43.3 tokens per second for the M5 and 28.1 tokens per second for the M3. And when I say tokens per second, I mean, like, think about a token as a word. So John might be a token. Stuart might be a token. It's not exact. I'm just making it simple here. So the M5 was able to produce 43 of these tokens or words a second, whereas the M3 is a bit slower, at 28.

4:00 Why? And it's not the GPU cores. We already covered that. This is a different metric. Okay. We're about to get a little bit nerdy here. So here's the reason why. The GPU can only produce or process one token at a time. So John and then Stuart and then Mills and then on. Now the math to produce each one of these tokens is actually pretty easy for the GPU. It's not like before where it had to process the entire prompt all at once.

4:23 That was a lot of math. Here, it's one token at a time. Easy math. Two plus two. Okay. I have to draw this. This Qwen model is roughly 15 gigabytes in size. That's the size on disk of its parameters or weights. Now here's where it hurts. Every time the GPU calculates a token, it has to pull all 15 gigs of that model out of memory to do the math for every token. Every time, 15 gigs. Now, again, the math to do this token or to create this token, easy for the GPU.

4:52 But now we're at the limit of what's happening right here, Memory transfer speed. The time it takes to get the 15 gigs of weights and parameters out of memory to the GPU so it can do its math. So why do you think the M5 was faster? Memory bandwidth. Looking back at our specs, 1.2 terabytes per second versus 819 gigabytes per second. That's a 50% increase. And looking back at our scoreboard, that adds up 1.5 times faster.

5:16 As you're watching this, you want one, don't you? Or at least you wanna dip your toe in a local AI somehow. Right? You do. The best place to do that, it's Micro Center. Micro Center is the perfect sponsor for me because I I go there all the time. I seriously do. Just to walk around. I don't even wanna buy anything sometimes when I'm there. I just wanna look at it, and I always end up buying something.

5:40 But for this video, I have to focus here. Their AI stuff, they got it all. You want a Mac Studio, which I think the new ones, the M5s, are available tomorrow, September 22. So go get you one if you're impressed by this thing. Otherwise, get another one. Do you want a NVIDIA DGX Spark? I have one of those. Get one of those. The AMD things, pretty much anything local AI, they're gonna have your back for. So if you're ready to dip your toe into this world and get control of your AI, go to Micro Center.

6:10 And also get whatever else your heart's desire because they have literally everything, Raspberry Pis, gaming PCs, TVs, Ethernet cables, Ethernet tools to build cables, just everything. UniFi. If you're if you have a home lab, you haven't been to Micro Center, go. They're actually opening up a new store in Austin soon. When did they say? It's opening this year. So Austin, look out for that. But if you're here in Dallas, dude, just go to one. And they have a bunch everywhere.

6:38 List right here. Oh, and also the Austin store. You can sign up for a free 128 gigabyte flash drive and redeem it when the store opens. Link below. So, yeah, that's awesome. So thanks again for Micro Center, my favorite store, for sponsoring this video. Back to the M5. Okay. With that in mind, let's try some different model sizes. Instead of a 27B, let's try a 14B. 32K? Let's start. The wait was only 14 seconds for the M5.

7:03 57 seconds for the M3. I was gonna say not too bad, it's kinda bad. Compared to the M5 Ultra, I mean, come on. Looking at our scoreboard, we're kinda widening the margins here. Four times faster on the wait or TTFT and 1.7 times faster on speed. Now, again, all we changed was the model size, and I wanna kinda hit that home because the larger model you go, the more stuff, the more memory and speed and GPU you need.

7:29 This 14B model is roughly eight gigs in size. That's 14,000,000,000 parameters. So the memory is transferring eight gigs each time. Okay. What about the 35B? Or the 122B? Now you might be wondering, Chuck, those numbers don't make any sense. They do. At first glance, I agree with you. But those models were not all the same types of models.

8:01 Some were dense. Some were a mixture of experts. I got the full test below and explanations as to what that is, but we gotta move on. Now it's time for testing transcription. Now just to make sure that you are aware that I do indeed add this Mac, here's my normal m three Mac. Let me go to about. Because I realized all I showed you was an empty box. But there it, there's my M3, 512 gigabytes of memory, and here's my M5, 256 gigabytes of memory.

8:30 I know some of you in the comments were thinking that. Now I transcribe with AI a lot. I transcribe my videos and my voice notes. What you're seeing here are gonna be a test between the M3 Ultra and the M5 Ultra with Whisper models from OpenAI. Those are local. You can run those locally. Okay. We'll start with a 13.8-minute recording. I don't know why it's 13.8 minutes. It's kinda weird. And we'll do it on the large-v3 model, and we'll keep going down.

8:52 Ready, set, go. Notice we do have stats running at the the entire time. M5 finishes. Did a pretty dang good job. 1.8 times faster. Let's go down a model. Let's go turbo. Start. Look at that. 2.3. Medium. Keeping in mind, the smaller we go on the model, the faster it can go or transcribe. 1.6 times faster. Let's go small. Man. Base, 2.1. And then tiny. Man. Alright. Now let's go for a larger recording. A whole day, which is two and a half hours.

9:22 No. Two hours and five minutes. I don't know why it's saying whole day. A whole day worth of recording. Okay. Let's go large-v3. Ready, set, go. Now pay close attention to the power usage. This is a killer thing about these Mac Studios, especially when you treat them like AI servers. You're not gonna get this running GPU farms or GPU clusters. 89.4 seconds for two hours and five minutes of audio on Whisper large-v3 . It's pretty good.

9:48 Verdict's in, 1.7 times faster. Let's go turbo. I won't do the rest because I'm getting bored. We'll see how fast turbo goes. 24.4 for two hours of audio. That's amazing. 2.3 times faster on turbo. Man, M5 Ultra is killing it. I'm gonna try tinies to play with that. Twelve seconds. Would I use this? A 100%. Like, transcription's one of those things that we can trust AI with. Open source AI does a great job. Open models do a great job with that.

10:18 And this is so fast. Two hours like that. And I transcribe footage, like, every single day. Now another thing I do as a YouTube creator is I have AI analyze my footage. Like, I actually look at my footage and tell me what's happening in that frame. So, when I have a given clip, I know what that clip's about. Like, if I'm filming me playing with a new Duo iPhone Duo, I want it to be able to analyze and go, yeah, Chuck was playing on a some kind of folding iPhone device.

10:45 Currently, I do this with the Gemini video API. Let's see how the new M5 Ultra will handle it, also versus the M3. Actually, let me play that file. I'm gonna go find out one of these Macs here so we can see how accurate it might be. Is it just me sitting here? Okay. Sure. It's doing one frame per second, and it's running the Qwen3-VL 32B 4-bit MLX. Let's start this. Ready, set, go. Dang. Okay. 22 seconds.

11:10 A man with a full dark beard and short dark hair styled upward at the front. Tattooing his left forearm that reads, let's just go. Okay. It did a really good job and very fast. So the M5 Ultra was 2.6 times faster with this, and the descriptions seem pretty similar. I mean, same model. Right? I actually might use this. I would even realize I'm sitting in a secret lab chair. Yeah. I'm paying for the Gemini video API.

11:35 This did a pretty dang good job. You know what I wanna do? This has been a project I've been kinda sitting on for a while. I wanna analyze home video footage, like old VHS style home video footage and, like, classify it, categorize it. Don't wanna pay for the hours and hours of footage of my mom, like, videotaping my soccer games. Right? But I would deal with this. That's a good idea. Yeah. I'm doing that. Anyways, next test.

11:58 Now it's time for images. The FLUX.2 klein 4B model for image generation. Now for many of you, may not have thought you could actually generate images locally. Let's see how good it's gonna be. What will be my prompt? A bald man running through a forest holding an iPhone Duo. We'll see if it knows. It it's not going to. But if it does, how cool would that be? And screaming at the top of his lungs as turtles are chasing him.

12:28 The turtles are wearing sombreros. Not sure if I spelled that right. Let's see how this goes and how fast will it be. They're pretty good. 2.6 seconds, 10.2, four times faster. Can I change the style? Make it photo realistic. I'm not sure if the tool I programmed here will do that, but we'll double check. Oh my gosh. It's so sick. No iPhone Duo, though. And by the way, the way the reason they look so similar is they're using the same prompt and the same seed.

12:56 And if you know what that is, you'll appreciate that. Oh, he's got the iPhone doing it. We did it, guys. Okay. Now let's go to clip. Now for video generation, we're gonna use that same photo from our last picture. I'll do a small model. What is this gonna do here? Woah. Not bad. Right? That's so fun. M5 was 1.7 times faster. And by the way, the model being used here is the LTX-2.5 22B model. It's a distilled model through the MLX port.

13:26 Let's try a bigger version. Start. That looks awesome. Wait. Is there sound? I don't know if there is. 2.5 times faster on a a longer clip. I'm honestly kind of amazed at how well it generated local video and images. I'm I'm, like, amazed at two things. The model itself, because I haven't tried that model, and also the speed at which the Macs created that. That's so fun, and it's free. You can do it endlessly. I gotta try one more thing.

13:54 I'm gonna describe something. A monster emerges out of a coffee cup full of black coffee. It roars in victory as it drinks all of the coffee. I gave it, like, a very little direction, but still, dang. So cool. Now what about Hermes? Can we use them with Hermes? Can we use the Mac M5 with local models and use a local model as our Hermes agent brain?

14:27 Let's try it out. Alright. Right now, I have launched Hermes. It's kinda freaking out on me with the M5 as the back end. Hello. It's kinda being weird. It's thinking. It's got a 131,000 context window. Hey. It actually responds. Dang. Okay. So what's actually happening here? I I didn't expect this. It's using the context for one of my main agents. It loaded that, and it's using Qwen from the M5. That's a lot of context. That's crazy.

14:57 Who am I? And what is my wife's name? That's a local model. Are you kidding me? Okay. Okay, Apple. Alright. It's pretty good. It's pretty good. I like it. I currently would not use that as my main model for my Hermes agent. However, as a tertiary agent for, like, basic stuff, maybe. I was honestly impressed by that. It loaded the the appropriate context, and that was impressive. That was yeah. I've never had good experiences with, local models in context.

15:28 I'm still floored by that. Yeah. I might I have to think of use cases. Comment below with what you think I should do or you do something. You tell me what you did. Time for the last test. Can it Jev with open source Kev or Laya? Let's try it. And if you don't know what I'm talking about, Jev is this new model made by TypeSafe what was that? TypeSafe AI. My my heck was coming out of me.

15:51 It's unlike ChatGPT or the local models we see here where you talk to them and talks back to you. You put in words and it spits out words to you. No. That's not what this one does. You put in words and it spits out binary answers or multiple choice answers. It's weird. It's crazy. Look it up. Jev has been taking the world by storm, but there are open source options. Local models, let's try them out. I have a Laya tab, which is one of the most popular ones right now.

16:15 It's gonna run a Jev like decision model on 300 labeled questions. I think it's, are these texts spam or not? Let's start. Essentially, it's showing it. Text after text after text after text saying spam, not spam, spam, not spam. Boom. 23 seconds, 300 questions answered, and it got it right. 100 of the texts were regular banking questions. 200 were spam. They both got it right, but the M5 Ultra was 1.4 times faster. And finally, I wanna do one last thing.

16:45 I wanna see if this local model can edit for me. If you have not seen this yet, DaVinci Resolve unlocked its editing application. It has an MCP with APIs that an AI agent can use to edit. I wanna see if a local model can interface with that and edit a video. Let's try it out. Okay. I have no idea if this is going to work. I've got my chat staged. This is, by the way, Bionic from LM Studio.

17:10 It's their new thing. I'm running Qwen3.8 27B. It's got the DaVinci Resolve MCP connected. I'm gonna give it this prompt. Now this is the system prompt system prompt. I kinda gave it just now, and let's have it do some editing. Go. I gave it auto review access. I can't believe this is all local. Okay. It looks like he's making some progress, and it's taking a long time. But he identified the cuts, and he's about to make some cuts.

17:33 Okay. He said he just finished it. Where? Is it here? Oh, wait. Where's the other project? He said, what? What's going on? Oh, there it is. Okay. So it looks like he did indeed edit it. Yeah. Look at that. It used DaVinci Resolve. That's pretty cool. Okay. Time to land this plane. The M5 Ultra is pretty sick. A lot of the things I do every day, I'll probably end up using this for when I get one because this is not mine.

18:04 Remember, Apple loaned this to me. Transcription, video generation, image generation, general local model use, like, it's really solid. Of course, in my AI stack, there will always be a place for frontier models for most of the work. But I can see myself giving more and more things away to the M5. I mean, the video editing was impressive. I'm probably not gonna use it for that purpose, but I think what we're seeing is a glimpse of what is gonna be coming down the pipe, what can be made possible.

18:29 Because like that video editing, that took when did I ask? Oh, it worked for thirteen minutes just on the on the editing, but I asked it it took a minute. What I cannot wait to do though is cluster some M5s. So coming out in October, they're gonna have the 512 edition. So the current one I have now has 256 gigs of RAM unified memory. The next one will be 512. And I would love to get my hands on four of those bad boys and cluster them together and run some serious local models.

19:01 Models that very they they do rival frontier models, and I wanna see what they can do. Now I just realized that out of all the testing I've been doing today, I didn't do any coding, and that's kinda funny. That didn't even occur to me as I was making this video and thinking about what I do because, like, that was weird. Anyways, that's all I got. Let me know what you think of the M5 below. And do you want one?

19:22 What's the price on this? I forget. M5 Ultra price. So the ultra chip, 80 core GPU, 256 gigs of RAM storage. Mine has eight terabytes of storage. Gosh. Storage prices are crazy. So for what I have right now, this is $15,000. Oh, $14,000. That's crazy. That's absolutely crazy. So knowing that, would you get one? $14,000, but it's unlimited AI.

19:53 It's also a pretty killer machine to do whatever you want with it. Right? I mean, it's the it's probably the best computer money can buy right now. It's not just an AI server, but I'm thinking AI only right now. That's where my brain's at. It's unlimited AI. I'd buy it. It would hurt because it's easier to digest paying frontier model prices, but you don't own it. This, when the AI companies go crazy, you still always have this, and that's kind of exciting.

20:18 Anyways, that's all I got. I'll catch you guys next time. Hey, you're still here. At the end of my videos, I like to pray for you, my audience. I know. It's kinda weird. Just go with it. One, two, three, pray. God, thank you for the person on the other side of the screen, on the other side of this camera. I thank you for who they are. I thank you that they are here. They're not here by accident.

20:40 So I just pray over their lives right now. If they're dealing with a hard time, if they're dealing with frustration, or just depression or anything that's just hanging on to them right now, a sin or just something that is dragging them down and robbing them of life and joy. God, I pray that you deliver that from them right now, that you reveal your truth to them, that you would give them a way out, show them the light, God.

21:13 And, ultimately, God, I pray that they know you, that, Jesus, they will know you, that they will hear your voice, and that you will shine a light in their heart, that you'll fill them with joy and everlasting peace, God. I thank you for this person, and I thank you that they're sitting here praying with me. It's so cool that we can do this. Bless them and their families right now, God. I ask in your name.

21:44 Amen. Hey. Thank you for letting me do that. I know it was weird, but it's also kinda cool. Right? I'll catch you next time.