Four 192GB Mini PCs Can Run DeepSeek-R1 671B in 2U. That Still Isn’t a 768GB GPU.
I think mini PCs have officially crossed into weird territory.
A year or two ago, the whole appeal was simple. You got a small box for office work, a home lab, maybe some light gaming, and you saved desk space.
Now Minisforum is showing four mini workstations in a 2U rack with 768GB of total unified memory, and the company says that setup ran a DeepSeek-R1 671B Q4 model locally.
Yes, 671B.
On mini PCs.
The model file in Minisforum’s test was around 380GB at Q4_0, and the company says a four-unit MS-S1 MAX-P495 cluster successfully ran it. Two units are also claimed to run Qwen3.5–397B at 16 tokens per second.
That sounds like a server story.
The machines are still roughly the size of thick books.
But there is a catch that matters a lot: four 192GB computers do not become one giant 768GB GPU just because you put them in the same rack.
That difference is the whole article.
Access without medium partner: 192GB Mini PCs Run DeepSeek-R1 671B
One box already has 192GB
The machine is called the Minisforum MS-S1 MAX-P495.
It uses AMD’s Ryzen AI Max+ PRO 495, a 16-core Zen 5 chip with Radeon 8065S integrated graphics and 40 RDNA 3.5 compute units. AMD’s NPU is rated at up to 55 TOPS, while Minisforum quotes 131 TOPS for the whole system.
The part local-AI people will care about is memory.
192GB of LPDDR5X-8533 unified memory.
Minisforum says as much as 160GB can be allocated as graphics memory.
That one number already puts the machine into a strange category. An RTX 5090 has 32GB of dedicated VRAM. This little box can expose up to five times that amount to its integrated GPU, although the RTX 5090 has far more raw GPU compute and much higher memory bandwidth.
Capacity and speed are not the same thing.
We keep relearning this with local AI.
A fast GPU is great until the model does not fit.
Four of them makes the headline stupidly large
Now multiply the memory by four.
Four systems at 192GB each gives you:
768GB of total unified memory.
If each machine allocates up to 160GB to graphics, that is as much as 640GB of graphics-addressable memory across the four nodes.
Across the cluster.
That last part matters.
The memory is physically attached to four different Ryzen systems. There is no magical switch that turns those four memory pools into one local HBM stack.
A distributed inference framework has to split the model across machines, move data between them, keep the layers or tensors coordinated, and deal with network delays.
So when somebody says “768GB AI cluster,” the number is real.
But it is not the same thing as owning one accelerator with 768GB of directly attached memory.
Not even close.
Minisforum’s 671B result is the part that got my attention
Minisforum’s product page goes beyond saying the system could scale.
The company says it tested it.
A four-unit cluster reportedly ran DeepSeek-R1 671B Q4, using a model around 380GB in Q4_0 quantization.
That is a very big model for something built from compact PCs.
The official page does not give a token-per-second number for that four-node DeepSeek run, which I wish it did. “It ran” and “it was pleasant to use” are two different statements.
I have been burned by this wording enough times now.
A model loading successfully is one milestone.
A model generating at 2 tokens per second is technically local inference too.
For the two-node test, Minisforum gives a much more useful number. It says two 192GB units ran Qwen3.5–397B locally at 16 tok/s.
That is vendor testing, not an independent benchmark.
Still, 16 tok/s is fast enough to be interesting. You can actually use that.
The 2U part may matter more than the 671B headline
This is the bit I almost skipped.
Minisforum says the machines are designed so multiple units can be placed into a 2U rack, with four-unit cluster support.
That changes who this is for.
A single 192GB mini PC is an enthusiast workstation.
Four of them in a rack starts looking like small-company infrastructure.
A research team could keep a private model inside the office.
A software company could put one cluster in a lab.
A media team could run internal retrieval and transcription jobs without sending files to a cloud API.
A university group could test huge quantized models without buying a full GPU server.
I am not saying this replaces a proper AI server. It doesn’t.
But the physical gap between “desktop AI PC” and “server” is getting blurry.
This thing is 222.1 × 206.3 × 77.1 mm and weighs 2.8kg per unit.
Four of them are still small enough to make the whole setup look a bit ridiculous beside the model size it is trying to run.
The network is where the fantasy meets physics
Distributed inference sounds easy when written as an equation.
192GB × 4 = 768GB.
Done.
Real life is annoying.
A large transformer model split across nodes needs to send intermediate data between machines. The exact amount depends on how the runtime partitions the model and how the model itself is built.
That traffic introduces latency.
Every time one node has to wait for another node, your expensive little cluster is sitting there doing nothing.
Minisforum gives the MS-S1 MAX-P495 dual 10GbE, plus two USB4 v2 ports at up to 80Gbps and two regular USB4 ports at 40Gbps.
Those are useful ports.
But the company’s cluster material does not clearly explain the exact interconnect topology used for every published model test.
That is something I want to see before treating the benchmark as settled.
Was the four-node DeepSeek run using 10GbE?
Was USB4 v2 part of the path?
What distributed runtime was used?
How were layers split?
What was actual output speed after the model loaded?
The official page tells us the result happened. It does not yet tell us enough about how.
And that is the kind of detail that can change a distributed LLM benchmark by a lot.
This is why 768GB is both true and misleading
Imagine four houses next to each other.
Each house has 192 liters of water.
The street has 768 liters in total.
But if your shower is connected to only one house, you do not suddenly have one 768-liter tank.
You need pipes between them.
Bad analogy maybe, but it gets the point across.
Distributed memory works because software knows how to divide a workload across the nodes. The model can be sharded so different pieces live on different machines.
That lets a model bigger than one machine’s memory run.
But every boundary between nodes has a cost.
Inside one chip, data can move through a very fast local memory fabric.
Across systems, it hits an external link.
That is why a four-node mini-PC cluster can beat one mini PC on capacity and still lose badly to a server with fewer total gigabytes but much faster GPU-to-GPU links.
More memory does not erase the network.
The Ryzen AI Max+ PRO 495 is built for this exact memory trick
AMD’s chip is what makes these machines possible.
The Ryzen AI Max+ PRO 495 has a 256-bit LPDDR5X memory interface and supports up to 192GB of unified memory.
That is very different from a normal desktop CPU connected to two DDR5 sticks while a separate GPU has its own fixed VRAM.
CPU and GPU share the same pool.
For local AI, that avoids one of the ugliest desktop problems: having 128GB of system RAM sitting beside a 16GB GPU that cannot fit the model.
The downside is bandwidth.
LPDDR5X-8533 over a 256-bit bus gives this platform a lot of bandwidth for an integrated design, but it is nowhere near the memory bandwidth of the fastest discrete GPUs.
Again: capacity versus speed.
The P495 wins the “can I load it?” fight much more often than a normal consumer GPU.
It does not automatically win the “how fast does it run?” fight.
Minisforum gives it 120W sustained power
This is another number that keeps the machine grounded.
The MS-S1 MAX-P495 can run the APU at 120W in Performance Mode, 95W in Balanced Mode, or 60W in Quiet Mode.
Minisforum quotes 160W peak power for the compute platform.
So four units could mean roughly 480W of sustained APU power if all four are running at the 120W setting, before you count the rest of the system losses and supporting hardware.
That is not tiny.
It is also not a rack full of multi-kilowatt GPU servers.
This is where I think the product gets interesting for smaller teams.
Maybe you do not need maximum tokens per second.
Maybe you need a huge model to fit, stay private, and run all day without turning one room into a furnace.
There is a real niche there.
And it has two 10GbE ports because Minisforum knows this is not a normal mini PC
The I/O is almost comical for the size.
You get two 10GbE RJ45 ports.
Two 80Gbps USB4 v2 ports.
Two 40Gbps USB4 ports.
HDMI 2.1.
Two M.2 slots with up to 16TB total storage support.
There is also a full-length PCIe x16 physical slot, although it runs electrically at PCIe 4.0 x4.
That slot is another example of why spec sheets need reading carefully.
“Full-length PCIe x16” sounds like you can drop in any giant accelerator and call it a day.
The electrical connection is x4.
And there are already user reports around the older MS-S1 MAX showing that external dedicated GPU compatibility can be messy depending on card, connection method and BIOS.
So I would buy this for the integrated 8065S and giant shared-memory pool first.
Treat expansion as bonus until somebody tests your exact setup.
192GB is soldered, obviously
There is no way around this.
The memory is LPDDR5X.
It is fixed.
You are not buying a 64GB model today and adding another 128GB next year.
That is the price of getting a wide, fast unified-memory interface into a machine this small.
Framework has the same problem with its Ryzen AI Max systems.
Apple has the same basic reality with unified memory.
Once local AI started caring about 128GB and 192GB shared pools, ordinary socketed RAM stopped being an easy answer.
It still bothers me.
A machine aimed at long-term workstation use would be nicer if memory were upgradeable.
But I also understand why these designs look the way they do.
You get the bandwidth by soldering the memory close to the processor.
Pick your poison.
The price is still missing
As of September 25, Minisforum’s US product page still literally shows:
$????
Then:
Price Reveal Coming Soon.
I kind of love that.
It is the most honest part of the page.
Because price decides whether this is a strange bargain or a tiny luxury server.
The previous 128GB-class systems were already expensive.
A 192GB LPDDR5X configuration, new PRO 495 chip, dual 10GbE, 320W internal power design and all this cooling will not be cheap.
And if you want four of them, multiply the pain by four.
That can change the comparison completely.
A four-node mini-PC cluster only makes sense if the total system cost stays well below the server alternatives you are trying to avoid.
Until Minisforum publishes pricing, I would not call this cheap local AI.
I would call it compact local AI.
Very different claim.
A 671B model also does not mean 671B quality for every job
Big parameter numbers look great in headlines.
They are getting less useful as a simple measure.
Mixture-of-Experts models may have hundreds of billions of total parameters while activating far fewer for each token.
Quantization changes memory use.
Architecture changes speed.
Context length changes memory pressure.
Draft models can change generation speed.
Cache design matters.
The runtime matters too.
So the fact this cluster ran a 671B model does not mean it will beat a smaller 120B model at your coding task.
It means the hardware crossed a capacity boundary that used to be reserved for much bigger machines.
That is still worth writing about.
Just do not turn parameter count into a benchmark score.
I think the two-node test is more useful than the four-node flex
The four-machine DeepSeek result gets the headline.
The two-unit Qwen3.5–397B at 16 tok/s result is what I would actually study.
Two boxes means 384GB total unified memory.
That is already enough room for models that are out of reach for almost every consumer desktop.
And two nodes reduce the networking mess compared with four.
If Minisforum can reproduce 16 tok/s across independent tests, with sane power use and stable long sessions, that setup could make sense for a small lab.
Four nodes are fun.
Two might be practical.
I did not expect to write that sentence about mini PCs.
This is also a warning for NVIDIA
NVIDIA still has the software advantage.
CUDA is everywhere.
A lot of AI tooling is written for NVIDIA first.
Its GPUs are much faster for many workloads.
That part has not changed.
But NVIDIA keeps making consumer cards with limited VRAM compared with these new unified-memory systems.
The RTX 5090 has 32GB.
AMD mini workstations are now hitting 192GB per node.
One AMD box can expose up to 160GB to graphics.
Four boxes can hold 768GB of total memory.
If your model needs 100GB just to load, it does not matter that the 5090 would be faster after loading it.
It never gets to that part without offloading or multiple GPUs.
That is the crack AMD and mini-PC makers are pushing into.
They do not need to beat NVIDIA at every benchmark.
They need to make models fit that NVIDIA’s consumer cards cannot hold.
The mini PC has become a memory appliance
This category is changing fast.
At first, mini PCs competed on size.
Then CPU performance.
Then integrated graphics.
Now local AI is turning them into memory appliances.
The CPU is still important.
The GPU is still important.
But the headline feature is becoming: how much model can this tiny box hold?
128GB was weird last year.
192GB now looks normal at the high end.
Four-node racks are already here.
This is happening much faster than I expected.
And there is one funny side effect: the mini PC is starting to make the full-size desktop look less ambitious, not more.
A giant ATX tower with 64GB RAM and a 16GB GPU can look under-equipped beside a 2.8kg box with 192GB unified memory.
Different workloads, yes.
Still funny.
I would wait for three things before buying one
I want independent inference benchmarks.
Not only a successful model load. I want output speed, prompt processing speed, power at the wall, thermals after an hour and performance across several quantizations.
I want the full cluster setup documented.
Show the cables.
Show the network topology.
Name the runtime.
Show exactly how the model is split.
And I want the price.
Especially the price.
As of today, Minisforum has shown enough to make the MS-S1 MAX-P495 interesting, but not enough to tell me whether four of them are a clever private-AI cluster or an expensive science project.
That answer could change overnight when the store page stops saying $????.
Four mini PCs running DeepSeek-R1 671B is the part I did not see coming
A few years ago, a mini PC was the machine you bought because you did not need a workstation.
Now Minisforum is building one that can be racked four at a time and used to run a 380GB quantized DeepSeek-R1 671B model.
That is a weird reversal.
The most interesting thing is not that four units give you 768GB of total memory.
It is that a company now expects enough people to want this that it designed a 2U deployment around mini PCs.
I still would not call it a 768GB GPU.
It is four computers.
That distinction matters for speed, software and network overhead.
But the model does not care what the boxes look like.
If the runtime can split 380GB across them and generate usable output, the old line between “mini PC” and “AI server” is getting pretty hard to see.
And Minisforum has not even told us the price yet.
That may be the scariest benchmark of all.
Sources I checked
Minisforum MS-S1 MAX-P495 official product page
Minisforum launch post: 192GB RAM for Local AI