Friday, June 12th 2026

Pre-Orders for $4000 AMD Ryzen AI Halo Mini PC Dev Kits Go Live

AMD announced in May 2026 that the Ryzen AI Halo Developer Platform mini PC would arrive in June, and launch day has apparently arrived, with the Strix Halo-powered mini PC now available for pre-order via Micro Center in the US. The Ryzen AI Halo developer kit is mostly aimed at AI developers, and although there are two SKUs, both have identical hardware, featuring the AMD Ryzen AI Max+ 395 CPU, with its Radeon 8060S iGPU, and 128 GB of LPDDR5x-8000 memory shared across both the CPU and GPU cores. The mini PC is launching with 2 TB of SSD storage, and it includes a 120 W PSU.

Pricing is set at $3,999, and the only difference between the two available SKUs is that one ships with Linux, while the other ships with Windows 11 Pro. No shipment date has yet been confirmed. Physical I/O is limited to USB Type-C, with three unspecified USB Type-C ports, likely USB 3.2 Gen 2×2 or USB4, one of which has DisplayPort Alt Mode support, an HDMI 2.1 port, a 10 Gbps Ethernet port, and a USB Type-C port for power input. It will also ship with both Bluetooth 5.4 and Wi-Fi 7. The mini PC itself measures just 149 × 149 × 43.18 mm, and the chassis appears to be mostly aluminium.
Source: Micro Center
Add your own comment

47 Comments on Pre-Orders for $4000 AMD Ryzen AI Halo Mini PC Dev Kits Go Live

#26
Red Hood
Wonder how much the dram cartel charge for that 128GB LPDDR5X-8000 RAM

The "ai" companies will buy this mini pc if they think it's worth it compared to what else is available to buy/rent that can run their models
Posted on Reply
#27
evernessince
adilazimdegilxSame spec mini pc is 3300$ from GMKtec and you can buy it right now. Why would you even want this instead?
Btw, based on the pricing between 64 and 128GB versions, 128GB RAM alone costs 2500$... So everything else is 800$...
I like the Chinese Mini-PCs but sometimes they can be a crapshoot. I tried miniforum machines and got 3 duds. 1 of those fell outside the amazon return window and minisforum support simply doesn't exist. 2 weeks for a reply and it was likely AI generated. Many horror stories regarding this. Zero issues with AOSTAR though, but I haven't had to contact their support. The machines actually work.

I can understand someone paying the extra to avoid the hassle, ultimately the downtime and time spent diagnosing issues may be worth it to some to just pay more.
Posted on Reply
#28
AusWolf
evernessinceI like the Chinese Mini-PCs but sometimes they can be a crapshoot. I tried miniforum machines and got 3 duds. 1 of those fell outside the amazon return window and minisforum support simply doesn't exist. 2 weeks for a reply and it was likely AI generated. Many horror stories regarding this.
That's bad news - I've been thinking about buying one if the Steam Machine ends up being too expensive. :(
Posted on Reply
#29
Tek-Check
Hecate91hallucinates completely wrong answers
Hallucination - sensory output that has no factual basis, e.g. hearing more than one voice in one's head, seeing angels, etc.
Confabulation - output based on filling gaps in memories with non-factual data or fabricating narratives

LLMs cannot hallucinate, as those models do not possess or process sensory information, like humans and other animals do.
LLMs can only confabulate, and that's exactly what they do when they mix-up facts, spew incorrect information, etc.

Visual AI models, such as Stable Diffusion, could hallucinate, as they produce sensory output, i.e. images.
Posted on Reply
#30
adilazimdegilx
evernessinceI like the Chinese Mini-PCs but sometimes they can be a crapshoot. I tried miniforum machines and got 3 duds. 1 of those fell outside the amazon return window and minisforum support simply doesn't exist. 2 weeks for a reply and it was likely AI generated. Many horror stories regarding this. Zero issues with AOSTAR though, but I haven't had to contact their support. The machines actually work.

I can understand someone paying the extra to avoid the hassle, ultimately the downtime and time spent diagnosing issues may be worth it to some to just pay more.
While I agree to that with almost any other part, these chips are pretty much same in every computer. There isnt many things you can do to customize it other than cooling, AIO and maybe some software. But even if you dont want a chinese brand, there is an Asus Z13 Tablet/Laptop with the same 128GB RAM that sells for 2850$ on newegg for example. There is also the ProArt laptop version with same specs in bestbuy for 3000$. While not exactly the same and probably more power/thermal limited at stock settings, I doubt you'd see 1200$ worth of difference in AI performance especially if you put these on some good laptop cooler and tweak it a little.

Considering you can even find full laptops with 3K OLEDs for 3000$, this AMD version shouldnt cost even 1$ more than 3000$ for a mini pc IMO. Even in this economy.
Posted on Reply
#31
kapqa
fi the new nvidia spoprts 1000tops, and these are about 128 tops or so, are those newer chips that much for usefuly in the AI dominion league?
Posted on Reply
#33
vigor
Lots of "geniuses" going to pay this price
Posted on Reply
#34
SamuelL
kapqafi the new nvidia spoprts 1000tops, and these are about 128 tops or so, are those newer chips that much for usefuly in the AI dominion league?
TOPS is a worthless metric, it’s just a vaguely defined number for supposed/theoretical raw throughput. It ignores various bottlenecks (including bandwidth) and doesn’t consider percision or quality.

The only useful measure of LLM inference is an identical workload, at identical precision, and checking tokens per second. TOPs = imaginary marketing bullshit.
Posted on Reply
#35
zenlessyank
No cartridge slot. No media drive. No included joysticks. What a crappy console.
Posted on Reply
#36
wheresmycar
All the consumer memory, SSD and GPU massive price hikes and so far in return we get a $4000 fancy AI dev shoe box with led lighting running along the bottom? From some of the comments I'm picking up on, i take it its basically consumer platform level performance masked as some sort of $4000 AI performance triumph?

Am I right in thinking "Unified memory" is the only real achievement here? But beyond AI models being memory constrained and Unified memory coming to the rescue, is the rest of the system basically just conventional hardware with normal performance features?

Just trying to wrap my head around these fancy shoe boxes. It hasn't interested me one bit to catch up with the weekly AI tech digest but recently a friend of mine came up with a pretty neat idea for developing a project management AI tool and its got me interested. Although, rather than AI, I prefer CA (computational automation). I'm still on the fence with surrendering to 'advanced information processing' as a form of independent thinking.
Posted on Reply
#37
evernessince
adilazimdegilxWhile I agree to that with almost any other part, these chips are pretty much same in every computer. There isnt many things you can do to customize it other than cooling, AIO and maybe some software. But even if you dont want a chinese brand, there is an Asus Z13 Tablet/Laptop with the same 128GB RAM that sells for 2850$ on newegg for example. There is also the ProArt laptop version with same specs in bestbuy for 3000$. While not exactly the same and probably more power/thermal limited at stock settings, I doubt you'd see 1200$ worth of difference in AI performance especially if you put these on some good laptop cooler and tweak it a little.

Considering you can even find full laptops with 3K OLEDs for 3000$, this AMD version shouldnt cost even 1$ more than 3000$ for a mini pc IMO. Even in this economy.
Yeah, if those other options are providing the same level of hardware and performance I don't see why you wouldn't go with them when you also get a good warranty.

I can't speak as to how well the laptop fairs compared to the dev kit, it'll be slower but almost certainly not to the degree that the price difference makes sense.

I wouldn't say that all products with the same specs are buitl equally. The motherboard and power delivery is very important. The quality of the traces, especially as interconnect speeds increase, are important as well. You don't want data errors and SSD dropping randomly. BIOS updates are important as it can cause various compat issues. My Gem10s for example don't like some SSDs (which took me awhile to diagnose, I thought my OS SSD was going bad when I was first setting the system up as it was stuttering and it took ages to find out it was just that specific SSD).

A big chunk of the cost increase is that it's a dev kit. It comes with it's own custom software stack designed to make using AI easy. Personally I don't think it's worth it and I don't think they should be upcharging this much.
AusWolfThat's bad news - I've been thinking about buying one if the Steam Machine ends up being too expensive. :(
My advice, if the steam machine idea falls though, look at reviews throughout the web on the product and company. There are some bad models and good models. Some companies have better customer service than others. You might also be able to snag a mini PC from ASUS, ASRock, or other big brand for a decent price as well.

Couple of additional caveats that many mini-PCs have is custom OS images and custom drivers. My minisforum UM870s required manual driver installs, off the shelf drivers wouldn't work. Required me to use their custom OS as well. My Gem10s did not, works with a stock windows ISO and the drivers that need custom installers (like the PSP) I don't bother with. This is why I place these machines on a vlan by themselves.

Even if the steam machine's price is bad, valve has an amazing warranty service, their products are polished, and you don't have to install sketch OS images and drivers.
Posted on Reply
#38
DeltaHawk001
I have a question:

What are dev kits for???
They all seem to be insanely expensive and seem to not have enough GPU power to train AI.
Posted on Reply
#39
Let3r
DeltaHawk001I have a question:

What are dev kits for???
They all seem to be insanely expensive and seem to not have enough GPU power to train AI.
Yeah. This is just AMD trying to cash in more on Strix Halo.
Posted on Reply
#40
AusWolf
evernessinceMy advice, if the steam machine idea falls though, look at reviews throughout the web on the product and company. There are some bad models and good models. Some companies have better customer service than others. You might also be able to snag a mini PC from ASUS, ASRock, or other big brand for a decent price as well.

Couple of additional caveats that many mini-PCs have is custom OS images and custom drivers. My minisforum UM870s required manual driver installs, off the shelf drivers wouldn't work. Required me to use their custom OS as well. My Gem10s did not, works with a stock windows ISO and the drivers that need custom installers (like the PSP) I don't bother with. This is why I place these machines on a vlan by themselves.
Oh I'm not going to have Windows on it, that's for sure. I'm planning either a custom SteamOS image, or just plain Fedora KDE.

The other system I'm looking at is the Arctic Senza, which also should come with decent warranty (I hope). But even that's only a plan B in case the Steam Machine falls flat on its face with pricing.
evernessinceEven if the steam machine's price is bad, valve has an amazing warranty service, their products are polished, and you don't have to install sketch OS images and drivers.
That much is true. Another reason why I'm still waiting.
Posted on Reply
#41
Xajel
Will only love a future upgraded version after the AI boom, when prices are actually good. I won't buy this or RTX Spark for that amount, not even half of it
Posted on Reply
#42
Grimling
do not expect big AI performance from this, the performance is limited by the LPDDR5x-8000 memory. So 10-20 tokens/s is realistic.
Posted on Reply
#43
Dahita
The thing I don't get is, Unified RAM ≠ VRAM. Sure, you fit bigger models, but it's awfully slow. The Nvidia Spark is the prime example.

A $4000 computer with a AMD AI PRO r9700 will have x4 the token output, for 32Gb of VRAM, and a much faster CPU for everything else.
Posted on Reply
#44
Legacy-ZA
Now you know where all the RAM has gone. :D

Spark and this, this, thing!
Posted on Reply
#45
Joshreddevil
LabRat 891Something nearly this small with the same specs was $1,299-1,999 less than a year ago... :(

Nice RTX Spark x86_64-copycat concept but-
only way I see something like this from AMD being competitive (+1ing for le$$) would be designing an RDNA4/5-equipped Zen5/6 APU around the least in-demand still-contractable HBM, and sticking it in a similarly-small box.
Nice that it has user upgradeable storage though, and open source drivers and software instead of paying $1000 a seat for NVIDIA AI Enterprise software suite,
Posted on Reply
#46
Vayra86
SamuelLTOPS is a worthless metric, it’s just a vaguely defined number for supposed/theoretical raw throughput. It ignores various bottlenecks (including bandwidth) and doesn’t consider percision or quality.

The only useful measure of LLM inference is an identical workload, at identical precision, and checking tokens per second. TOPs = imaginary marketing bullshit.
Smells remarkably like TEN GIGARAYS!

DeltaHawk001I have a question:

What are dev kits for???
They all seem to be insanely expensive and seem to not have enough GPU power to train AI.
So called creators and prosumers that are really just attic nerds with too much cash and no real skillset

Apparently there's enough of those to make up a market.
Posted on Reply
#47
Grimling
DeltaHawk001I have a question:

What are dev kits for???
They all seem to be insanely expensive and seem to not have enough GPU power to train AI.
A Hardware Development Kit (HDK) basically is a Piece of Hardware that you buy to develop, in this case, AMDs Profit!
The more you buy, the more you develop!
Posted on Reply
Add your own comment
Nov 21st, 2026 06:11 PST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts