Strix 128GB "Deal" $3300

[FrgMstr]

I have watched this Strix based agent computers going hugely up in price the last few months. A buddy did send over the deal he found. If you are wanting to give local AI a try, this might be your cheapest option in the 128GB unified memory footprint left out there. No idea on this builder

https://www.amazon.com/ACEMAGIC-M1A-126tops-Graphics-Pro/dp/B0H2HN59VJ

ACEMAGIC M1A PRO+AI Mini PC, AMD Ryzen AI Max+ 395 (120W 126tops, 5.1GHz), Radeon 8060s Graphics, 128GB LPDDR5x 8000MHz, 2TB PCIe 4.0 SSD, WiFi 7, Mini Gaming PC Windows 11 Pro | UP 5.1GHz Radeon 8060s Graphics, 128GB LPDDR5x 8000MHz, 2TB PCIe 4.0 SSD, WiFi 7, Mini Gaming PC Windows 11 Pro Mini Computer​

​

 

[Supercharged_Z06]

I thought ordering a GMKtec Evo 2 back in early June of 2025 for $1800 was a bit nuts... It seemed kinda pricey just for a beefy mini PC at the time... Zero regrets now. Currently sells for $3649

 

[Spun Ducky]

I thought ordering a GMKtec Evo 2 back in early June of 2025 for $1800 was a bit nuts... It seemed kinda pricey just for a beefy mini PC at the time... Zero regrets now. Currently sells for $3649

All the regrets not buying several of the frameworks on preorder. This point my jaunt into this ecosystem will be gorgon halo 192gb or medusa halo whenever that is out.

 

[FrgMstr]

All the regrets not buying several of the frameworks on preorder. This point my jaunt into this ecosystem will be gorgon halo 192gb or medusa halo whenever that is out.

I got my last Framework on "sale." I checked a couple days ago, exact same spec was over $1100 more than I paid for mine.

My son bought a Framework about a month ago. He was like, "Shit, Dad, that is a lot of money." I replied with, "It's going to get worse, before it gets worser." He popped.

 

[StoleMyOwnCar]

If you look at the Microcenter open box section, there are some 128GB laptops with the same chip in them for less than 3k USD. Obvious caveat is you need a Microcenter near you, but it's a nice way to get a travel-capable PC that you can also game on while running these models.

On the other hand, I'm not sure if I would go for these over the DGX Spark, personally. It's just a faster chip, and they have a port in them which allows you to join together multiple Sparks for more inference power and VRAM later, if you want to expand. They are more expensive on a per-unit basis. On the other hand I got my Spark for around 3k, and the current prices are brutal.

Edit: For example:

Obviously YMMV.

I don't think a desktop form factor in this case would provide too much more cooling or power so it's more attractive. I almost got one of these.

 

[FrgMstr]

If you look at the Microcenter open box section, there are some 128GB laptops with the same chip in them for less than 3k USD. Obvious caveat is you need a Microcenter near you, but it's a nice way to get a travel-capable PC that you can also game on while running these models.

On the other hand, I'm not sure if I would go for these over the DGX Spark, personally. It's just a faster chip, and they have a port in them which allows you to join together multiple Sparks for more inference power and VRAM later, if you want to expand. They are more expensive on a per-unit basis. On the other hand I got my Spark for around 3k, and the current prices are brutal.

Spark is a solid deal if you want to stack nodes, no questions asked. Spark is faster, but not terribly. Spark wins out on prefill bigtime, not always on the token gen however....just tied to a NV CPU and Cuda in a lot of aspects. The way I look at this is no one is buying one of these and expecting frontier perf. Once you get your model and agents built out solidly, who cares if it is 40TPS or 50TPS. You are going to be asleep.

 

[Spun Ducky]

I got my last Framework on "sale." I checked a couple days ago, exact same spec was over $1100 more than I paid for mine.

My son bought a Framework about a month ago. He was like, "Shit, Dad, that is a lot of money." I replied with, "It's going to get worse, before it gets worser." He popped.

Besides expanded ram in the next versions that tempt me, I agree with your later post that as the price of strix halo catches the spark the spark is looking like the better deal for other reasons. Prefill as you mentioned and then its nvidia, its as close as plug n play exists right now for local ai. This point its a lot of seeing a "good" deal and just pulling the trigger. I have gpu compute but that stuff runs so hot. I think we got at least 2 more years of crazy prices before it slows down or at least used market catches up as substantial jumps happen in this small box systems.

I am becoming far more interesting in long horizon thinking from a big slower model and then let little models run wild. So the while a sleep statement is spot on.

 

[FrgMstr]

I am becoming far more interesting in long horizon thinking from a big slower model and then let little models run wild. So the while a sleep statement is spot on.

Local AI for SMB or private, this is your tradeoff in the current ecosystem. Buy the memory footprint, give up some TPS. Buy the GPU, give up unified memory footprint. Not a lot of middle ground right now in terms dollars spent.

That said, there are a lot of good models that can fit inside a 32/16GB VRam buffer. But I have not seen one that can pull off what Qwen 35B can at Q8 or Laguna at 4Q. Real screen pull from right now. I will trade fast for better results.
|

 

[StoleMyOwnCar]

Spark is a solid deal if you want to stack nodes, no questions asked. Spark is faster, but not terribly. Spark wins out on prefill bigtime, not always on the token gen however....just tied to a NV CPU and Cuda in a lot of aspects. The way I look at this is no one is buying one of these and expecting frontier perf. Once you get your model and agents built out solidly, who cares if it is 40TPS or 50TPS. You are going to be asleep.

It has a lot of active development because the idea is that you can use spark cluster(s) as a development platform before you can deploy the exact same thing straight to enterprise level Cuda hardware (6000 Pro, etc). For sort of close-to-frontier model running, it takes about 2 sparks connected by 1 cable. This can run DSV4 Flash at native quants with 1 million token context, which people have found to be very capable overall. That's the only thing that is tempting me to go for another spark. I guess my point is moreso that the option is there if you choose the Spark. If you just need a one time solution and you don't mind speeds, the AMD platform is definitely easier. Especially if you would like a laptop anyway, which is why I sort of recommend the laptop. The only challenge might be getting it to boot headlessly if you install Linux so that you can conserve some VRAM for VLLM/etc, since it kind of comes with a screen attached. Spark is Ubuntu-native, which means it's as simple as setting up SSH access and then connecting to it via LAN IP.

You could also try Qwen 3.5 122B. I'm not sure if there is a quant of it that can run as efficiently on that platform but might be worth trying.

 

[mnewxcv]

If you look at the Microcenter open box section, there are some 128GB laptops with the same chip in them for less than 3k USD. Obvious caveat is you need a Microcenter near you, but it's a nice way to get a travel-capable PC that you can also game on while running these models.

On the other hand, I'm not sure if I would go for these over the DGX Spark, personally. It's just a faster chip, and they have a port in them which allows you to join together multiple Sparks for more inference power and VRAM later, if you want to expand. They are more expensive on a per-unit basis. On the other hand I got my Spark for around 3k, and the current prices are brutal.

Edit: For example:

View attachment 818377

Obviously YMMV.

I don't think a desktop form factor in this case would provide too much more cooling or power so it's more attractive. I almost got one of these.

The rog flow z13 is not a traditional laptop form factor either. It borders on being a tablet and has a "kick stand". Deal breaker for some.

Strix halo is a great platform. The eye watering price from a year ago is now a price we wish we could pay today. It's funny how we went from looking at 128gb as overkill just a little while ago, and now wishing it had more.

For what I've played around with on ai and LLMs etc, 128gb is around the sweet spot for using it as an ai appliance, while the upcoming 192gb configs sound ideal for being able to do all the ai stuff in the background while still doing everyrhing a workstation needs to do with the remaining ram. Currently, running a large LLM and trying to video edit at the same time runs the system out of memory unless I compromise something.

Lastly, this is a fantastic platform, ai aside. I have a framework desktop mainboard with the basic 8 core 32gb config and it's a little beast. Great steam machine alternative.

 

[StoleMyOwnCar]

The rog flow z13 is not a traditional laptop form factor either. It borders on being a tablet and has a "kick stand". Deal breaker for some.

Strix halo is a great platform. The eye watering price from a year ago is now a price we wish we could pay today. It's funny how we went from looking at 128gb as overkill just a little while ago, and now wishing it had more.

For what I've played around with on ai and LLMs etc, 128gb is around the sweet spot for using it as an ai appliance, while the upcoming 192gb configs sound ideal for being able to do all the ai stuff in the background while still doing everyrhing a workstation needs to do with the remaining ram. Currently, running a large LLM and trying to video edit at the same time runs the system out of memory unless I compromise something.

Lastly, this is a fantastic platform, ai aside. I have a framework desktop mainboard with the basic 8 core 32gb config and it's a little beast. Great steam machine alternative.

The one I actually tried out at some point was not the Flow but rather an HP variant that was a traditional laptop (nice sturdy metal frame, too) with an OLED display. It was about the same price or a little cheaper at the time. Unfortunately my screen had some issues and by that point I was already intent on moving onto the Spark. I'm glad I did back then.

As far as VRAM, I think you will quickly find that if your needs ever grow, the amount of capital you can dump into this is possibly massive. The Spark is actually kind of a bargain, considering what it can do in a cluster, vs any other solution cost wise, when you consider what it can host. Obviously at this point they're all 4k minimum, so you're looking at 8k for even two of them, though. If you were trying to do that with RTX Pros, you would be looking at 22k, which is the price of about 5 sparks.

It's definitely not a perfect platform, but for just pure LLM it's hard to beat from a cost, space, and energy efficiency standpoint at the moment. If you do want an all around machine instead, then yes the AMD variants are fine.

I would be highly dubious about the 192GB config prices. If they end up costing significantly more, it's kind of a wash at best. Unfortunately they don't come with Linux, either, so you have to install Linux (harder that it sounds; driver support for these in my experience is not the best when I tried it on the HP laptop; freezing browsers and all kinds of issues) if you want to get all your VRAM ready to use immediately.

 

[FrgMstr]

Those Halo boxes come with Windows and Linux installs. And setting up WSL and Ubuntu could not be simpler.

 

[FRZ]

the Asus ProArt GoPro edition is also worth looking into over the Flow. Has 128GB RAM, costs $3k, and comes in a traditional laptop form factor.

 

[StoleMyOwnCar]

If they come with dual boot out of the box, then that's good. I had a lot of issues with trying to run Mint in any stable fashion on the HP laptop. Note that Windows tries to randomly search and destroy any Linux partitions it finds sometimes IME, so yeah. Careful with that.

 

[mnewxcv]

the Asus ProArt GoPro edition is also worth looking into over the Flow. Has 128GB RAM, costs $3k, and comes in a traditional laptop form factor.

I got one of those today. So far so good. An advantage of the hp is 2280 nvme ssd vs the asus being 2230. The HP has great build quality but looks like an HP so won't get a second glance, while the proart gopro looks pretty unique / enthusiast and I bet it would draw a bit more attention. Still, neither one screams 16 cores and 128gb just looking at it, they just look like regular laptops. That part is still pretty unbelievable. I think the years of 10lb desktop replacements are finally gone.

 

[FRZ]

I got one of those today. So far so good. An advantage of the hp is 2280 nvme ssd vs the asus being 2230. The HP has great build quality but looks like an HP so won't get a second glance, while the proart gopro looks pretty unique / enthusiast and I bet it would draw a bit more attention. Still, neither one screams 16 cores and 128gb just looking at it, they just look like regular laptops. That part is still pretty unbelievable. I think the years of 10lb desktop replacements are finally gone.

I got one recently too. Its definetely slick looking without being too ostentatious. With RAM prices what they are, makes the $3k I paid for the proart seem like a great deal, especially with the potent 8060s iGPU

One gripe I will say I have with the machine is the 60hz screen. Not sure what asus was thinking not going 120hz for a machine of this caliber

Excited to put it to use and start messing around with LLMs

 

[mnewxcv]

Got the px13 with ai max 388/32gb against the px13 gopro with ai max 395/128gb. Going to see how much you can do with 32gb, and also see how much 16 vs 8 cores matters.

 

[myzz]

No idea on this builder

google Acemagic spyware... credential stealers, malicious chrome extensions, some reports of the malware remaining after reinstalling windows...

I'd definitely avoid this brand.