SerpApi.Org
A $3,999 mini PC with no discrete graphics card in it is a strange thing to get excited about for AI, except for one thing: you can load a model that won't fit anywhere else. That's the whole pitch for AMD's Ryzen AI Halo, the Strix Halo dev box built on the Ryzen AI Max+ 395 with 128GB of unified memory. It'll hold gpt-oss-120B, Qwen 3 Coder Next, or even DeepSeek V4 Flash with a special quantized version. In fact, it'll load a lot of models that a 32GB RTX 5090 can't even begin to load, and it'll run many of the Mixture of Experts models at a speed you can actually live with.
I've spent a lot of time with a 128GB unified-memory machine, though it isn't this one. It's the Lenovo ThinkStation PGX, Nvidia's GB10 Grace Blackwell box, and working with it has taught me what a shared pool of memory can do for you while also highlighting the imperfections that would be easy to overlook at first glance. The Ryzen AI Halo is AMD's take on the same idea, at a cost that's now cheaper per gigabyte all while relying on ROCm rather than CUDA. Still, the trade underneath is identical; you get an amount of memory a graphics card will never offer, and you give up the bandwidth and the software comfort that make a graphics card feel fast.
Unfortunately, if you just bought one, you may not be happy to learn that, for the overwhelming majority of what people do with local models, I'd still go for a used RTX 3090 that costs significantly less than a Halo does. The Halo wins on one comparative metric, and it wins it convincingly, but it isn't the fight most people actually really care about.
The Ryzen AI Halo holds models a graphics card can only dream of
Capacity is the entire goal
The Ryzen AI Halo has 128GB of unified memory, and it's shared between the CPU and the integrated Radeon 8060S as one coherent pool, and up to 96GB of that can be handed to the GPU as VRAM. You won't have to deal with weights transmitting across a PCIe bus, nor do you need to split a model across two or more cards. You don't need to try and figure out weird ways to get a model to fit in memory; at this size, chances are, it just does.
Every graphics card has a hard wall: an RTX 5090 tops out at 32GB, a used RTX 3090 or RTX 4090 gives you 24GB, but if a model's weights and KV cache go over that, speed drops off a cliff. Alternatively, the model may refuse to load at all. The Halo doesn't hit that wall until you're looking at models that are bigger than 96GB of VRAM, and at this size, that's a different class of problem entirely.
These devices have "slow" memory, clocking in at 256GB/s, but that's not so much of an issue these days. A lot of the most popular models are mixture-of-experts designs now, meaning the model keeps an enormous number of total parameters resident but only reads a small number of them, namely the active experts, for each token it generates. gpt-oss-120B holds all 120 billion parameters in memory yet activates a few billion per token, so the bandwidth you need to feed it is far lower than the raw size implies.
There's a catch, though, as this only works for MoE models. If you load a big dense model, like Devstral 2 123B, where every parameter fires on every token, it'll slow to a crawl. That's because now it really does have to read all of the weights for each token generated, and the bandwidth just isn't there. In other words, the Halo isn't a machine that runs big models well, it's a machine that runs big MoE models well.
I've used a 128GB unified memory machine, and I know its limits
Memory bandwidth is a constraint
How big of a model you can actually fit in is usually the headline associated with devices like these, but it's the half of the picture that actually flatters unified memory in the first place. When generating a response using any model, generating a reply is constrained by memory-bandwidth, which these machines respond to well when the model in use is an MoE model. But before a model can generate anything, it has to read your prompt, and reading the prompt, known as prefill, is compute-bound rather than bandwidth-bound.
The problem here is that an integrated GPU carries a fraction of the raw compute a discrete card does. So the longer your prompt, the longer you sit watching nothing happen before the first token even arrives. On short back-and-forth chats you'll probably never notice, so long as your system prompt isn't all that long. However, give it any kind of long input, like a document, a big codebase, a big block of RAG context, and the wait before the model starts talking will be far longer than anything someone on a dedicated GPU will experience.
I know how that feels because the PGX does exactly this, though Nvidia's software stack and Blackwell compute helps keep prefill times competitive enough. Even then, long context still feels sluggish compared to a proper GPU sitting in a desktop. The Halo runs the same kind of workload with weaker compute and ROCm underneath instead of CUDA, so I'd expect the wait to be at least as long, not shorter.
None of this makes the machine bad, it makes it a machine with a very specific usecase. If your local AI life is mostly short prompts and long generations, you'll rarely be bothered by that prefill ceiling. However, if you push the enormous context that made you want 128GB in the first place, though, you'll see it everytime you send a message.
For anything that fits in 24GB, a used 3090 is much better
And a lot of models do
A used RTX 3090 has 24GB of VRAM and has a memory bandwidth of, roughly, 936GB/s. The Halo has far more memory but has a bandwidth of around 256GB/s. For any model that fits inside 24GB, that bandwidth gap is a pretty big deal, and the older card wins in a head to head easily. On top of that, fitting a model inside 24GB of VRAM is a lot more doable than you'd think. It's every 7B, every 14B, and even a lot of 32B models when quantized, and those sizes are where the good, everyday-usable local models actually live. Most people never need to run anything larger.
For example, I run Qwen 3.6 27B on my AMD Radeon RX 7900 XTX, and that card has 24 GB of VRAM. It's fast, especially with MTP, and I can have 100,000 tokens of context. The prefill is slower as I'm using Vulkan for generation instead of ROCm (Vulkan has faster generation speeds, and that's the trade-off you make), but it's still incredibly capable. Even if I did outgrow that card, the answer would still be another GPU for more memory. Two RTX 3090s would get you to 48GB, and could actually make a 70B model possible to self-host locally. Oh, and it would still cost less than the price of the Halo, while still keeping all the speed the single card had on anything smaller.
That particular card isn't without its issues, to be clear. It launched in 2020, it draws 350W and will noticeably heat your room while it's generating tokens, and the used market is still quite expensive. Still, for raw capability per dollar spent, it's still the best value in local AI, and it isn't close.
What all of this means is simple: what you're really buying the Halo for isn't speed. On everything the 3090 can hold, the 3090 is faster, usually by a lot. Instead, you're buying the ability to hold models in memory that the 3090 can't, and that's a far narrower usecase than it seems on the surface.
A graphics card is more than a box that hosts one model
Games, videos, and more
Putting these AI machines against a dedicated GPU often ignores that a graphics card does a hundred other jobs, while also doing most of them just... better. When it comes to image generation with SDXL or Flux, video generation, it's all written for CUDA first and everything else later. On a 3090 that will run with no fuss, but on the Halo it'll be painful, and not because the box is short on memory. SDXL barely troubles a 24GB card, never mind a 128GB one. It's painful because hardware outside of CUDA is less of a priority in general.
The problem is that the gfx1151 architecture in these chips is still officially in preview on AMD's own ROCm support list more than a year after the hardware shipped. To be fair to AMD, It has got a lot better, and on the right kernel it now largely just works. But the "right kernel" is important; use the wrong one, and you'll need to set environment-variable overrides so ROCm doesn't mistake the chip for a different GPU. For many of these things, it's possible to run them on AMD's hardware, but it'll take you longer to get it to that point than a CUDA user ever needs to spend on it to achieve the same result.
Then there's the other half of a graphics card's life that has nothing to do with AI. A 3090 can be put into a PC that you already own and plays games properly, which the Halo's integrated graphics can do, sure, but it won't match its performance.
Finally, a graphics card holds its value and can move to your next build when you upgrade. The Halo's memory is soldered LPDDR5X, so what you buy on day one is what you own forever, and it's one sealed unit. A GPU, though, is a component you keep reusing, whereas a mini PC is a single commitment you make once.
The Halo answers one question, but most people aren't asking it
You can hold a lot in memory
There's one advantage AMD has that can be hard to beat with any GPU, though. Even if you pair two RTX 3090s together in order to pool 48GB of VRAM, 48GB still won't hold gpt-oss-120B at a quant you'd actually want to use. if you want that, you're looking at a four-card rig or a used server accelerators, which means a lot more money, far more power, a louder machine. Oh, and that whole machine would have to be built around them, rather than being a self-contained box on a shelf drawing maybe a couple hundred watts of power maximum.
On capacity per watt and per square inch, nothing from the consumer world touches the likes of the Halo or GB10-class devices. If your goal really is to run 100-billion-parameter MoE models at home, and you've made peace with the prompt-processing speed that comes with the territory, it's about the cheapest and cleanest way anyone has to do that right now.
But that's a small audience of people who actually care. Most people who think they want to run a 120-billion-parameter model at home would be happier, and faster, running Qwen 3.6 27B on a card that costs a quarter as much. It'll be faster for everything that fits, a fraction of the price, image generation and gaming and the rest handled too, and it's still going to worth real money when you decide to move on.
The Halo is a specialist, and a good one at that. Nothing at the price matches it for the single thing it does: hold a lot of AI stuff in memory. If you're buying it, buy it for that reason alone and you'll be happy with it for years. Just be honest with yourself about whether that's the question you actually have. At $3,999, it's an expensive way to find out it wasn't.
