I've spent a bit of time testing this brand new mini PC, the Ser 10 from Beelink, [music] against last year's model, the Ser 9, and the Ser 8 before
that. They all look kind of alike, but inside [music] they're different because the Ser 10 has AMD's new Gorgon Point chip versus the previous Strix
Point. And after doing my developer and LLM AI tests, I came out with something I didn't expect. All
right, so this is the Beelink Ser 10 Max. They added Max to it. It's got the Ryzen AI 9 HX470, code name Gorgon Point. But the real shopping question
here is three machines, not one. [music] If you take a look at the Beelink site, last year's Ser 9 still available, still hot, and still on sale.
Usually a couple hundred bucks cheaper. In fact, I've been using mine now for a while, for more than a
year, as a media server. And as a side note, this is what it looks like if you just have it on a rack for over a year. It's got a little bit of dust,
but it's not [music] too bad. Really solid machine. The other option is the M 4 Pro Mac mini, which I also have set up here. I'm including that
because the M4, the base M4, caps out at 24 gigs, which is really not that great for a lot of the
workflows I'm going to show you. And also, since we're comparing it to [music] the Ser 10, which is a 32 and 64 gig configurations, then we should go
to the M4 Pro. For the M4 Pro though, Apple charges some serious money for that. This is the basic M4 Pro, not even the 20-core GPU, and it's 48
[music] gigs, not the 64-gig version. Did they even have the 64? They got rid of that? Let's go to 1
terabyte, 10 gigs ethernet, suddenly you are at $2,000, and you're only at 48 GB of RAM. So, the Beelink site still makes sense. You're only going up
by couple hundred bucks, and you're way under that. This does come with 1 TB of storage. So, the question isn't is Ser 10 a good upgrade? It's should
I get a Ser 10, Ser 9, or save up for the M4 Pro or the upcoming M5 Pro, whenever that comes. But,
this is a performance video, so we're going to get into that. So, these days I'm always flipping between models. GPT for research, Claude for coding,
Nano Banana for image generation, VEO Kling and Runway for video, six tabs, six bills, and counting. Enter Chat LLM Teams. One dashboard houses every
top LLM and route LLM picks the right [music] one. GPT Mini for ultra-fast answers, Claude Sonnet
for coding, Gemini Pro for massive context. They've recently added Gemini 3 and GPT 5.1 the moment they dropped. Create professional presentations
with graphs, charts, and deep research detailed content. Need human-sounding copy? Humanize rewrites text to defeat AI detectors. [music] Need
visuals? Pick frontier or open-source models. Nano Banana, Midjourney, and Flux for images, Magnific for
upscaling, plus VEO Wan and Sora for video. All built in. You also get Abacus AI deep agent to pretty much do anything. Build full-stack apps,
websites, reports with just text prompts and deploy them on the spot. They have Abacus AI desktop, which is the brand new coding editor and assistant
that lets you vibe code and build production-ready apps. And the kicker? It's just $10 a month, less than
one premium model. Head over to chat.abacus.ai or click the link below to level up with Chat LLM Teams. By the way, if you don't know what a Gorgon
is, this is a Gorgon. I don't know why AMD has to go with scary things for their [music] chip names. This one will turn you into rock if you make eye
contact, so don't do it. Some quick specs on this chip, we've got 12 cores and 24 threads, Zen 5
architecture plus Zen 5c, eight of those, and it boosts up to 5.2 GHz. Check it out. Shared memory, half of the memory can be shared, 29 GB in my
case. So, AMD's marketing material obviously will include the MPU. The use cases are limited, but I'll show you how we can use that later on in the
video. And the iGPU is here too. This happens to be the Radeon 890M. If that sounds familiar, you're
right. >> [laughter] >> It's the same one that was in the SER [music] 9. More on that later. The memory here, I have 64 GB on my machine, but it does
come with 32. And this is DDR5 5600, user configurable and user upgradeable, which is different than the SER 9 was. You can also upgrade the storage
up to 8 TB if you can find it and afford it. A few other life improvements like USB4, HDMI 2.1,
DisplayPort 1.4, and 4K at 240 Hz triple monitor support. And 10 gigabit network ethernet. So, that's pretty cool considering the SER 9 only had 2.5.
Now, if you're so inclined, there's also nice-looking orange one. Is it orange or red? I don't know, but it's got OpenClaw pre-installed on it, and
they're charging 100 bucks more for it. But, look at this option right here. 96 GB with 2 TB SSD for
[music] $2300. This 96 GB for that much is actually not bad right now. If I spec the 2 TB version of a Mac Mini, we're at 2500 bucks, and we're half
the RAM. But, it's kind of funny that they're charging 100 bucks more for the OpenClaw version of this. And it's definitely not as bad as that guy on
Craigslist who was charging $1500 for a base Mac Mini with OpenClaw pre-installed on it. All right,
let's start light. Speedometer 3 is a JavaScript benchmark. It simulates real web application interactions, the kind of thing you feel every time you
open a sluggish site, single-threaded, and everyone's got a frame of reference for this one. The SER 9 hit 31.5 in the seven-way comparison video that
I did not too long ago. That one had a bunch of other mini PCs included. Last year, the M4 Pro Mac
Mini got 45.2 in that. The base M4 got beat the M4 Pro, got 47.8, but they're close. Apple is kind of hard to beat in single core scores right now.
They've got that market cornered. [music] And we'll see that pattern again. Our speedometer is a JavaScript runtime, Web Tooling Benchmark is another
one that I use by V8. Yeah, those are the same V8 people that actually write the engine. This tested
JavaScript tool chain like TypeScript, Babel, Terser, and stuff like that. Here the Sir 9 came in at 20 on the TypeScript score and 18.8 on the
geometric mean. Both Mac Minis are way ahead, 35.99 each on TypeScript. By the way, the M4 base and the M4 Pro tie on that one. The geo mean is where
they separate a little bit. And after running the Sir 10 test, it landed on 34.14. That's what TypeScript
and 28.62 for the geo mean. Okay, hold on. Now, that's 65% jump on TypeScript over the Sir 9. 65. And the geo mean is up over 50% here, too. This is
by far the biggest generation-over-generation win I've seen so far for this machine. And look at that TypeScript score. It's within 5% of both the
Macs. By the way, I'm looking over here because that's where my computer with the chart is. So, the Sir
9 trailed the Macs by 75% on [music] this test last year. This year the gap is essentially gone. The geometric mean is still behind by 11 to 13%. Not
a clean tie there, but this mini PC has finally caught up to Apple's Mac Mini on a real V8 tool chain benchmark. All right, the obligatory Geekbench
benchmark. I don't usually like to dwell on this, but in this case, it's important. Everybody uses
Geekbench to measure this stuff, so might as well throw it in. And last generation Sir 9 came in at 2889 for a single core and 14890 for multi. The M2
Pro Mac Mini, that's the one from [music] the It's the bigger one, okay? From a couple years ago. That one was 2,700 and 13,000. Now, the M4 Pro,
that's the number to beat. It's got pretty decent score, 3945 for single and 15,321 for multi. You can
see the Sierra 8 was quite a bit behind on that one. And the Sierra 10, that lands at 2993 single and 15,216 multi. Pretty big leap over last
generation. And for the single-core, it's not a huge difference. However, the multi-core, compared to the M4 Pro Mac mini, that's a difference of less
than 1%. It basically ties Apple's flagship Mac mini on multi-thread Geekbench. Maybe this will translate
to when I do compilation tests in a bit. All right, it's time for the meat. Can't believe I just said that. That's just Some vegetarians might object,
but I don't care. I I eat meat. I like it. It's good. Lamb is my favorite. Beef, but a good steak is really good. All right. Python brought. >>
[laughter] >> This is a test that I run to show how pushing all the cores to the max with an interpreted
language like this, interpreted test like this. This is a Python algorithm, and it's the channel's signature move. It pegs every single core to 100%.
Oh, yeah. There it goes. You can see that working quite nicely there. Now, the Sierra 9 already beat the M4 base on this. 28.64 seconds versus [music]
31.41. Shorter number is better in this case. Be like ahead of Apple silicon here, folks. And
that's the previous generation. As an aside, I also ran this on the X Elite mini PC that came out Some of you might remember that. Well, the Sierra 9
tied with that machine, which is the first-generation X Elite that was the most powerful one. I want to see the X2 Elite in mini PCs. That's going to
be a killer one. Anyway, M4 Pro was still the one to beat, 23.24 seconds on that one for this test,
considerably faster. And the sear eight was quite a bit slower, 42.73 seconds on that one. So, we saw a 33% jump from sear eight to sear nine. That's
huge. How does the sear 10 do? Hmm, okay. Sear nine was 28.64. Sear 10 is 28.9. A nice number here. This is the fastest that it got. That's literally
a quarter of a second difference. So, basically, kind of the same. Still beats the M4 base. So, that
part of the story holds. But, generation to generation, nothing here. So, this was an interpreted multi-core test. But, before we get into compiled
multi-core test, I want to back up here and show you this project, which is something I ran before on those machines as well. This is a large mono
repo by Victor Savkin. This is NX mono repo. It includes 26,000 NX JavaScript components. There's a lot
of components, but they are very small. So, it corresponds to a medium-size enterprise repository. Basically, JavaScript at scale. There's five
Next.js apps here, 20 libraries, 250 components each. And NX provides a cache, but I ran this build without a cache, just to see what it's like. Now,
before I get your hopes up, this test didn't work out so well because uh there's repository drifts, and I
just don't believe the results cuz it can't be like this. Last year, sear nine, we got 14.1 seconds. The M4 Pro was 7.76. The M4 base was 7.36. Yeah,
I know. M4 base won again. Third test in a row. So, what I like about this test is it's a TypeScript tool chain. It runs through node. Lots of process
spawning. Lots of file system hits. Apple silicon's been good at this for years. So, this is where
I thought the sear 10 has to climb, right? Maybe not dominate, but climb. And sear 10 landed at 2.91 seconds. And I thought, come on. Five times
faster than the sear nine? I don't think so. So, because of the way this test is set up, the NX tool itself probably got faster and optimized and
[music] unfortunately, I'm going to have to say goodbye to this tool because um yeah, when you're doing
benchmarks, you can compare things side to side, side by side, over time, but when you're doing real-world projects, then things tend to drift a
little bit. So, it's not the same workload anymore. We can't really compare to the previous builds. However, it builds this thing in under 3 seconds,
which is pretty incredible. And that's good for both the chip, the machine, and the mono repo, and the
tool. So, good for them. We're going to skip this test next time, though. I want to move on to a benchmark that I designed. This is a .NET build, and
this build basically generates 100,000 namespaces and classes, each one with recursive calculations and nested loops, so the compiler can't just throw
them away and skip the work. It has to do it. Same workload every time, year after year. This is
the test where the Cier 9 beat them for base model, 91 seconds. That beat them for, which got 106.7 seconds. The Cier 8 was 109 seconds, and the only
machine that was meaningfully faster was the M4 Pro at 66.4 seconds. So, this is the test for Gorgon Point. If Zen 5's IPC story is real, and the math
behind the new generation actually shows up in compiled code, he's going to land the hardest here.
And this is also the loudest test. I'm hearing the fans here. Here it is being built. Pretty much all the CPU cores are involved in this one, although
they don't go as hard as that matter broad test, but still, it's more realistic. And we got 90.9. This is a difference of 1/10 of a second. This was
supposed to show the new chip's strengths. Basically, we got the same result as last time. But, the
Cier 10 still beats the base M4. So, Beelink continues to be a better choice than the base Mac mini for a compiler heavy .NET work. But, on the M4
Pro, Apple silicon leads still. Now again, the previous test was synthetic, so I wanted to go and to build a real thing, Umbraco. That's a open source
CMS, it's been around for ages. It's a mature .NET project, open source. But again, this is a real
project on GitHub. So again, there's going to be variation over time. It's constantly evolving, they're adding features, they're refactoring, pulling
in new dependencies. So the number I might get might not be fully repeatable year after year, but if you have a machine and you want to build it right
now compared to my result that I'm going to get, then at least it's going to be valuable in that
way. If you want to compile this yourself, you can find it on GitHub right over here. They're constantly updating it. 11 hours ago was the last
update. Can't believe this thing has been around so long. Sir 9, 149 seconds for the compilation. Sir 8, 161. So gen over gen on the Umbraco was only
7%. Why is that? Well, because real builds are IO heavy. There's disk, there's NuGet, the file system. And
again, Mac dominates this one. M4 Pro, 84 seconds. M4 base model even destroys the B link, 93.9 seconds there. All right, got it done. Sir 10 lands at
161 seconds. Now hold on, that's slightly slower than the Sir 9. What's up? And it just happened to match the Sir 8 from a couple years ago. Both 161
seconds. So again, on this one, a real .NET production app, full release pack, every project, the
Sir 9 happened to be the best of the three here at 149. By the way, I want to flag something real quick here. The first time I ran this, I got 217
seconds on the Sir 10. And that was with Windows Defender running. So um I had to turn that off. So if you're doing this kind of compilation, you got
Windows Defender running, you might want to exclude the your code directory from that. >> [music] >>
And the way I have my systems set up, first of all, I automated the build, the deployment of this package, and I do have videos on how to set up a
development environment on a Windows machine, and on Macs. So, check that out. It's on my channel. And I recently created a script to automate that
whole process, too. I'd recommend going through the whole thing just to know what's getting installed,
and then you can use the script on your subsequent builds if you want. It's up on my GitHub. I'll link to it down below, but I made a special video
for members of the channel. Thank you to the members, by the way, for supporting the channel. I made a video on how to run it. It's all in the
instructions on the GitHub, though. If you want to just go check it out, I'll link to it. All right, let's
switch topics to another popular thing that's happening these days in 2026. Oh, I don't know. What could that be? AI. AI. Nice. >> [laughter] >> This
thing has a GPU, a CPU, and an NPU. So, this little box has three chips on it, and they all can do AI. There's a 12-core CPU, Radeon 890M iGPU, and
there's a 55 tops XDNA 2 NPU. AMD says all three together do 86 tops of AI work. Tops. Yeah, we all
know the marketing stuff. Now, out of the box on Windows, the most popular tool people use to run local LLMs uses one of those, but it's the wrong
one. So, let's do it. I'm going to do Ollama run Qwen [music] 2.5 7B, and we'll do verbose, and I'll do my architecture prompt, which produces [music]
a pretty long output. There it is. Results are streaming. Looks pretty good, but let's check task
manager. The CPU looks pegged, but the GPU is not being used at all. NPU, zero. It doesn't even show up here. Three accelerators on this box, and
[music] one of them, the regular CPU, is doing all the AI work, and we don't want that. That's going to be slow. Well, there is Vulcan flag that you
can set on the environment for Ollama that will allow you to use the GPU. So, on day one, the AMD AI chip
you just bought is going [music] to be doing AI on the part that isn't the AI part, doing AI on the part that isn't the AI part. We got a result at
14.2 [music] tokens per second. So, to enable that on Windows, you just go to the environmental variables [music] and set the Ollama Vulcan flag to
true. If there isn't one, just add it. And then just make sure you restart Ollama. All right, here we
go. That looks like it's going a little faster, doesn't it? Let's take a look at the GPU, and yes, [music] there's the GPU compute. It's happening on
the GPU now, which is beautiful, 100% utilization there. So, I ran a couple of them as a test. Llama 3 0.23B, we went from 27 tokens per second to 37
and a half. Gwent 2.5 1.5B went from 47.5 to 68.3 tokens per second. So, there's some differences
there. So, we've got the iGPU thing going, but there's still one chip in this box we haven't even touched. And that's this >> [music] >> NPU. And
guess what? The old one had it, too. AMD's been putting this thing on the box for 2 years, and nobody actually benchmarks it. So, first try, I used >>
[music] >> Lemonade Server. They got these hybrid recipes where the NPU does the prefill and the iGPU
does the decode. Those are the two [music] stages of inference. It even comes with this little handy app. Let's do a high here. It's kind of hard to
see down here, but tokens per second is 14.3 on this one. This is a 1 billion parameter model. We can see that it bumps the NPU just a tad bit, and
then it uses the GPU for the decode. What if I use my longer prompt here? Boom. Okay. Let me see a
little bit more NPU activity. That's the prefill. And the prefill happens pretty quickly. Most of the work happens on the GPU, though, cuz that's the
decode phase. So, the NPU is kind of built for that stage, for the prefill. It's a different job, different job, [music] different chip. So, I ran a
longer prompt, and it was a 4400 token prompt with a short reply. And that's the actual prefill
workload. On Quen 7B, the CPU got 255 tokens per second there. The Vulcan iGPU 240 and the NPU with a hybrid mode 631 tokens per second. That's 2 and
1/2 times the iGPU on the same chip. So, if you're streaming chat, use the iGPU. If you're doing rag, agents or coding assistance, anything that dumps
a long context, then you want to use the NPU. And since this is a lemonade server, you can actually
connect it to a coding agent. It's for another video. Perhaps you want to see that, let me know in the comments down below if you do. I'm curious
myself cuz I haven't actually tested lemonade before on the channel. But now I want to show you something that this hardware does that if you just
read the specs, you'd say, "No way, that's not possible." Quen 2.5 14B, that's the Q4 quant, over 8 GB on
disk. Now, the iGPU on this chip, it shows you 4 gigs of memory, just 4. And I'm going to load 8.37 gigs of weights onto a 4 gig iGPU. This shouldn't
work. And it's loaded. All 49 layers on the GPU. How? All right, so I'm watching this in real time. Get counter on the right side, GPU adapter memory.
Before I loaded the model, it was 0.69 GB dedicated, 5.26 shared. And after, 3.62 dedicated, 17.36
shared. 20.98 GB of iGPU memory in use right now. On a chip, the device manager says there's only 4. So, that's UMA, unified memory architecture,
which the M4 [music] and the Apple silicon also has, but this has it, too. So, when the iGPU asks for more memory than its allocation, Windows just
quietly grabs some from system RAM and says, "Here you go." Hands it over and the model has no idea. So,
prompt processing was about 50 tokens per second, generation 8.8 tokens per second at 14 billion parameter model. And by the way, on this machine,
there's still like 25 more gigabytes of unified memory headroom left here. So, a 22 billion parameter model would fit. Probably even a 30 billion
parameter model would fit, too, with Q4 on the iGPU. Now, remember that the SER 10 and the SER 9 have the
exact same GPU in there. So, if you bought a SER 9 last year for the AI stuff, >> [music] >> your hardware is still fine. Update your software. So,
quick recommendations here. You should buy the SER 10 if you care about the MPU for rag and long context AI work. If you want upgradeable RAM because
SER 9 doesn't have that. Or if you want that 10 gig ethernet port. You should probably skip it if you
already have the SER 9 and your main workload is the iGPU LLM path. Save your money until the next generation, which is Medusa, comes out in 2027.
Another scary creature that I wouldn't want to mess with. Now, besides these, there's actually a bunch of other mini PCs that you can consider and I
made a video up here and up here. You can check those out. You [music] might be interested in those.
Thanks for watching and I'll see you next time.
[music]