Framework built its name around one very simple idea: if you buy a computer, you should be able to open it, fix it, replace parts, and keep using it.
That is why its new desktop is such a strange machine.
The upcoming Framework Desktop with AMD’s Ryzen AI Max+ PRO 495 packs 192GB of LPDDR5X unified memory into a chassis that is only 4.5 liters. It gets a 40-CU Radeon 8065S GPU, 273GB/s of memory bandwidth, a 16-core / 32-thread Zen 5 CPU, and up to 131 TOPS of total AI compute.
For local AI, those numbers are huge in a box this small.
And then you reach the memory.
All 192GB is soldered.
You cannot add another stick later.
You cannot buy the cheaper version today and move to 192GB in two years.
You cannot replace one failed memory module.
The memory configuration that makes this computer so interesting for local AI is also the part you have to choose forever.
That sounds completely against Framework’s whole identity.
The funny thing is: Framework knows that better than anyone.
Access without Medium Partner: Framework New AI Desktop
192GB in 4.5 liters is the real headline
Forget TOPS for a second.
The number that matters here is 192GB.
Local AI has turned memory capacity into something much more important than it was for normal desktop buyers.
An RTX 5090 has 32GB of GDDR7 VRAM.
A 64GB workstation sounds large until you try to load a heavily quantized 70B or 100B-class model with long context and other apps running.
The Framework Desktop goes the other direction.
AMD’s Ryzen AI Max+ PRO 495 supports 192GB of unified LPDDR5X-8533 memory through a wide 256-bit memory interface. AMD says the platform can expose as much as 160GB as dedicated graphics memory.
That is the kind of number that changes what can fit locally.
A model does not care that your GPU is fast if half of it has to spill into slower system memory.
This is why the Ryzen AI Max family has become such a weird local-AI favorite. It is not trying to beat a huge NVIDIA card at raw GPU speed. It is giving the integrated GPU access to a memory pool that consumer NVIDIA cards simply do not have.
192GB total. Up to 160GB for graphics.
That is the pitch.
And for some local models, capacity wins before speed even enters the conversation.
The Radeon 8065S is not an RTX 5090
This part needs to be said clearly because these memory numbers can get silly very fast.
The Framework Desktop’s Radeon 8065S has 40 RDNA 3.5 compute units.
That is useful.
It is not a 5090.
An RTX 5090 has much higher raw compute and vastly higher memory bandwidth. If the model fits in the 5090’s 32GB VRAM, NVIDIA can be much faster depending on the model, runtime and kernel support.
CUDA also remains the easier road for a lot of AI software.
So no, a Framework Desktop with 160GB available to the GPU is not secretly a five-times-better RTX 5090.
The comparison is about what fits, not only how fast the first token appears.
This has become the basic split in local AI:
NVIDIA gives you speed. These big unified-memory AMD boxes give you room.
Sometimes you want both.
Then the budget gets ugly.
Framework already knew soldered memory would annoy people
This is the part that makes the story more fair than the headline.
Framework did not discover soldered memory last week.
When it launched the first Framework Desktop in 2025, the company openly explained why Ryzen AI Max memory could not be modular.
Framework said it spent months working with AMD to see whether removable memory was possible. The problem was the processor’s 256-bit memory bus and the bandwidth it needed.
The original 128GB Framework Desktop used LPDDR5X-8000 and delivered around 256GB/s of memory bandwidth.
Framework’s conclusion was simple: it could not make removable RAM work at the required throughput.
So the memory was soldered.
The company did not hide that.
I still do not like it.
But there is a difference between “we soldered it because upgrades are bad for business” and “the electrical design of this chip needs a wide, high-speed LPDDR interface that normal DIMMs cannot provide.”
That difference matters.
The new 192GB version makes the compromise harder to ignore
At 32GB, soldered memory is annoying.
At 64GB, you think about it.
At 128GB, it starts becoming part of the buying decision.
At 192GB, the memory is basically the reason you are buying the machine.
That changes the emotional part of the purchase.
If you spend workstation money on a computer because it can run large local models, you probably expect that machine to stay useful for years.
But the memory is frozen on day one.
Buy 128GB and decide later that you need 192GB?
You are not buying RAM.
You are replacing the entire mainboard, including the processor and integrated GPU.
That is technically an upgrade path because Framework sells replaceable mainboards.
It is also a very expensive way to upgrade memory.
The old board is not useless. You could reuse it in another build or sell it, which is better than throwing away a complete proprietary computer.
Still.
Calling a motherboard swap a RAM upgrade is stretching the word a bit.
The rest of the machine is very Framework
This is why I do not think the soldered RAM makes the whole product hypocritical.
The Framework Desktop uses a standard Mini-ITX motherboard form factor.
It uses a FlexATX power supply.
The cooling fan is a standard 120mm unit.
Storage is replaceable through M.2 slots.
The case is designed around replaceable parts and swappable exterior tiles.
The system even has a PCIe slot.
On the new 192GB configuration, Framework is opening the physical end of that PCIe x4 slot, which means longer PCIe cards can fit even though the electrical connection remains x4.
Framework also says the new model will ship with a Noctua fan installed, and a prebuilt Fedora option is planned.
So this is not a sealed box where nearly everything depends on one proprietary enclosure.
A lot of the machine can still be repaired.
The two big exceptions are exactly the parts AMD integrated most tightly: the processor and the LPDDR5X memory.
That is the trade.
273GB/s is why removable DIMMs are not the obvious answer
People see 192GB and immediately ask:
Why not just use four DDR5 sticks?
Because that would be a different memory system.
The Ryzen AI Max+ PRO 495 uses LPDDR5X-8533 over a 256-bit interface. Framework lists 273GB/s of memory bandwidth for the upcoming configuration.
That bandwidth matters a lot for integrated graphics.
A discrete GPU has its own GDDR memory sitting beside it.
An integrated GPU normally has to share slower system memory with the CPU, which is one reason integrated graphics traditionally lose badly in high-end workloads.
Ryzen AI Max tries to solve that by giving the processor a very wide, fast unified memory interface.
The GPU gets access to a huge pool without crossing a PCIe link.
That is the entire trick.
Normal replaceable desktop memory would make the machine easier to upgrade, but it would also change the bandwidth and board design that make the Radeon 8065S interesting in the first place.
Framework chose performance and capacity.
Repairability lost one round.
This is probably the worst year possible to sell 192GB of soldered LPDDR5X
There is another problem.
Memory is expensive right now.
Framework has been unusually open about that too.
In September, the company said LPDDR5X costs were continuing to rise, forcing another price increase on the existing 128GB Framework Desktop.
Framework had already raised that configuration earlier in the year as cheaper memory inventory ran out.
That makes the upcoming 192GB model interesting for a second reason.
We still do not know its price.
As I write this on September 30, Framework’s official page still says “192GB coming soon.” Pre-orders are scheduled to open in the morning Pacific Time, but pricing has not been published yet.
That means the number everybody wants is still missing.
And it could decide the whole product.
The original 128GB Ryzen AI Max+ 395 Framework Desktop launched at $1,999 in 2025.
Framework has since had to adjust pricing as LPDDR5X costs climbed.
Now add another 64GB of faster LPDDR5X-8533, a newer Ryzen AI Max+ PRO 495, and a memory market that is much worse than it was at the original launch.
I would not assume this will be cheap.
GMKtec has already shown how ugly 192GB pricing can get
We have a useful comparison from this week.
GMKtec’s new EVO-X5 Pro also uses the Ryzen AI Max+ PRO 495 and offers 192GB LPDDR5X-8533.
Its 2TB configuration is listed at $6,799, while the 4TB version goes to $7,099. Very limited early-bird pricing drops that to roughly $6,399–$6,699.
That does not mean Framework will cost the same.
The products are configured differently, and Framework historically priced its desktop aggressively for the amount of unified memory it provided.
But it tells us something about the market.
192GB is not normal consumer-PC memory anymore.
It is becoming workstation pricing.
Framework’s final price will tell us whether its version is still the strange value pick or whether the whole 192GB mini-desktop category has moved into luxury territory.
I am waiting for that number more than the TOPS figure.
131 TOPS sounds nice. I care more about 160GB for the GPU.
AMD rates the Ryzen AI Max+ PRO 495 platform at up to 131 TOPS total AI compute.
The NPU alone is rated at up to 55 TOPS.
Those numbers look good on a product page.
For local LLM inference, they are not the first numbers I would use to decide whether to buy this machine.
I care about 192GB total unified memory, up to 160GB available as graphics memory, and 273GB/s memory bandwidth.
Then I care about real tokens per second in LM Studio, llama.cpp, Ollama and whatever runtime is best for this chip.
TOPS mixes different hardware blocks and precision levels into one big number.
A local model has much more specific needs.
Can it fit?
Can the runtime use the GPU correctly?
What quantization are you running?
How long is the context?
How fast is prompt processing?
How fast is token generation after the cache is warm?
Those are the numbers I want.
The 128GB Framework Desktop already gave us a clue
Framework has published local-AI results for its existing Ryzen AI Max+ 395 systems.
On the 128GB version, Framework lists OpenAI gpt-oss-120b at around 38 tokens per second using LM Studio on Fedora 42 with MXFP4.
That is the kind of benchmark that makes these machines interesting.
It is not about putting a small 8B model on expensive hardware and celebrating 200 tok/s.
The appeal is running models that normally push consumer GPUs into multi-card setups or CPU offload.
The 192GB model widens that door.
Not because 192GB automatically makes the same model faster.
Because it gives you room for larger models, larger context, multiple models, or several local agents at once.
That is a better use of the extra memory than chasing one benchmark number.
192GB also makes clustering more interesting
Framework has been talking about clusters since the original Desktop.
The existing machine has 5Gb Ethernet and two USB4 ports, and Framework says multiple systems or bare mainboards can be connected to run larger models through tools such as llama.cpp RPC.
The Mini-ITX board can also be installed in rackmount cases.
With the new 192GB system, two machines would give you 384GB of aggregate unified memory.
Notebookcheck also reports that Framework’s new open-ended PCIe slot makes it possible to install a 50GbE adapter and link systems using RDMA over Ethernet.
Again, aggregate memory is not one giant local memory pool.
Distributed inference has network overhead.
But 384GB across two tiny systems is a serious amount of local model capacity.
And now we are back to the same weird place mini PCs keep taking us in 2026.
A product that looks like a small desktop starts behaving like infrastructure.
The soldered RAM is more defensible here than in a normal laptop
I am usually much harsher on soldered RAM in ordinary PCs.
A thin office laptop with 16GB soldered forever?
No thanks.
A machine built around a 256-bit LPDDR5X interface feeding a 40-CU integrated GPU at 273GB/s?
I understand the reason.
I still want replacement options if the memory itself fails.
That is the part where repairability gets uncomfortable.
If one LPDDR package develops a fault outside warranty, the practical repair unit is the mainboard.
Framework is better positioned than most vendors because it actually sells mainboards and publishes repair information.
But replacing a board containing the CPU, GPU and 192GB of memory because one memory component failed is not the kind of repair story people normally associate with Framework.
There is no neat solution here.
The engineering that makes the machine good at local AI is the same engineering that makes the memory difficult to replace.
That is why this product is more interesting than another “Framework broke its promise” headline.
It did not.
The hardware problem is real.
The price will decide whether this is clever or ridiculous
Right now, Framework has given us nearly every headline number except the one that matters to buyers.
192GB.
273GB/s.
40 GPU compute units.
131 TOPS.
16 Zen 5 cores.
4.5 liters.
No price yet.
If Framework lands far below the $6,000-plus 192GB mini PCs appearing this month, this could become one of the most interesting local-AI desktops in the market.
If it lands close to them, then the comparison changes.
At that point you start looking at DGX Spark, used workstation GPUs, multi-GPU towers, Mac Studio configurations, or just paying for an API.
Local AI gets less romantic when the little box costs as much as a used car.
That is the part nobody should skip.
Framework built an upgradeable computer around a non-upgradeable idea
The Framework Desktop is still modular in ways most mini PCs are not.
The case is serviceable.
Storage is replaceable.
The fan is standard.
The power supply is standard.
The mainboard is replaceable.
The memory is not.
And the 192GB memory is the exact feature that turns this from a small desktop into a serious local-AI machine.
That contradiction is not going away.
I think Framework made the right engineering choice.
I also think buyers should understand what they are buying.
With normal desktop RAM, you can start small and grow later.
With this machine, memory is a day-one decision.
Choose 192GB because you need it, not because you think you might casually add it later.
You cannot.
The upside is that 192GB inside a 4.5L computer gives local-AI users a kind of capacity that used to require much larger and more expensive hardware.
The downside is printed on the motherboard.
No DIMM slots.
Just 192GB soldered in place, feeding a GPU that finally has enough memory to make that compromise worth arguing about.
Sources I checked
Framework Desktop — 192GB configuration
Framework: Introducing the Framework Desktop
Framework: Memory and storage pricing updates
Framework: Choosing a Framework Desktop for Local AI
AMD Ryzen AI Max+ PRO 495 specifications