Ryzen AI Max PRO 495 vs Max 395

The new 192GB mini PCs look like a huge generational jump when you read the spec sheet. Framework, GMKtec, Minisforum and other vendors are moving from AMD’s 128GB Ryzen AI Max+ 395 machines to systems built around the new Ryzen AI Max+ PRO 495, and the memory number alone makes the new boxes look like a different class of computer.

Then I compared the GPU. The old Ryzen AI Max+ 395 has a 40-CU Radeon 8060S running up to 2.9GHz, while the new Ryzen AI Max+ PRO 495 has a 40-CU Radeon 8065S running up to 3.0GHz. Same RDNA 3.5 generation, same 40 compute units, and only a small clock bump. AMD did not suddenly give these $7,000 machines a GPU twice as large.

The price changed far more than the graphics hardware did. Framework currently lists its 128GB Ryzen AI Max+ 395 Desktop at $3,449, while the new 192GB Ryzen AI Max+ PRO 495 DIY Desktop starts at $6,799. That is only 64GB more memory, but the total machine price is almost 97% higher.

This is the part I think matters for local AI buyers. The new 192GB machines are not mainly selling you a faster GPU. They are selling you the right to load models that the 128GB machines cannot comfortably hold.

Access without medium partner: 192GB Mini PCs Cost Twice as Much

The cleanest comparison is Framework: $3,449 versus $6,799

Framework makes this unusually easy to see because it sells both generations in almost the same 4.5-liter desktop format. Its current 128GB Ryzen AI Max+ 395 DIY configuration is $3,449, while the new 192GB Ryzen AI Max+ PRO 495 DIY version starts at $6,799.

That is a $3,350 jump. In percentage terms, the new machine costs about 97% more, while unified memory rises from 128GB to 192GB, which is a 50% increase.

Of course, this is not a pure RAM upgrade. The new system gets the newer PRO 495 processor, slightly higher CPU clocks, faster LPDDR5X-8533 memory, a Radeon 8065S instead of the 8060S, and some platform changes. Framework also ships the new version with a Noctua fan and has changed a few details around the board and expansion setup.

But Framework itself says the important local-AI change is the extra 64GB of unified memory. Its September 30 technical post says the CPU and GPU architectures remain the same between Strix Halo and Gorgon Halo, and that theoretical memory bandwidth increases by only about 6.7%.

That is a surprisingly useful admission. The expensive new model is mostly a capacity upgrade, not a massive speed upgrade.

The GPU changed from 40 CUs at 2.9GHz to 40 CUs at 3.0GHz

AMD’s own specifications make the contrast even clearer. The Ryzen AI Max+ 395 uses Radeon 8060S graphics with 40 compute units and a maximum graphics frequency of 2.9GHz.

The Ryzen AI Max+ PRO 495 uses Radeon 8065S graphics. It also has 40 compute units, and the maximum graphics frequency rises to 3.0GHz.

So the compute-unit count did not change. The graphics clock increased by around 3.4%.

There are other differences in the processor and memory system, and real workloads do not scale perfectly with clock speed anyway. Still, if you only looked at the GPU block, you would never guess the complete machine could cost almost twice as much.

That is why I would be careful with the phrase “next-generation AI mini PC.” The new boxes can run much larger local models, but the part doing the GPU math is not twice as large or twice as fast.

Memory bandwidth improves, but not by anything close to the price increase

The older Max+ 395 platform uses LPDDR5X-8000 over a 256-bit memory interface. AMD’s own Ryzen AI Halo reference system lists about 256GB/s of memory bandwidth for that setup.

The new PRO 495 raises memory speed to LPDDR5X-8533. Framework lists the new desktop at 273GB/s, which is where its roughly 6.7% theoretical bandwidth increase comes from.

That extra bandwidth is useful. Local LLM generation on integrated graphics can be heavily memory-bandwidth limited, so even a modest increase can help token generation and prompt processing.

But 6.7% more bandwidth is not a new performance class. The real new capability is that the memory ceiling moves from 128GB to 192GB.

That sounds obvious after you say it, but product pages make it easy to get distracted by newer chip names, TOPS ratings and AI branding. For local LLMs, the extra money is buying space first.

This is not a bad upgrade if your model needs 150GB

Here is where I think the criticism needs to stay fair. If the model you want requires 145GB, the 128GB machine is not “almost as good” as the 192GB one.

It simply does not fit cleanly. You can offload parts of a model to SSD or CPU memory in other architectures, split workloads across machines, use a smaller quantization, or choose another model. Every workaround changes speed, quality, complexity or all of them at once.

A single 192GB unified-memory system avoids that mess. Framework says the extra memory lets users keep larger models loaded on one machine instead of streaming weights from SSD or clustering systems, and it can also let you keep two smaller models resident at the same time.

For an agent setup, that second use case is more interesting than it sounds. One fast model can handle the main conversation while a larger, slower model stays loaded for difficult calls, and you avoid paying the load penalty every time the agent switches.

So yes, 192GB has real value. The weird part is how expensive that last 64GB has become.

GMKtec proves this is not only a Framework pricing problem

Framework is not the only vendor asking workstation money for 192GB. GMKtec launched the EVO-X5 Pro with the same Ryzen AI Max+ PRO 495 and 192GB LPDDR5X-8533, and its regular pricing is $6,799 with 2TB or $7,099 with 4TB.

The first 30 units were offered at $6,399, with a seven-day early-bird tier at $6,599. Even the aggressive launch price still puts the machine deep into high-end workstation territory.

GMKtec says the system can run fully offline LLM inference with models up to 320 billion parameters, depending on model, quantization, software and configuration. That is the sales argument for 192GB in one sentence: not “our GPU is dramatically faster,” but “we can hold a model that your smaller machine cannot.”

I actually think that is a more honest way to look at the market. These machines are becoming local-AI memory servers that happen to sit on a desk.

Minisforum makes the price gap even harder to ignore

Minisforum currently sells the older MS-S1 MAX 128GB Max AI Compute Edition with a Ryzen AI Max+ 395 for $3,799 on sale. Its new 192GB MS-S1 MAX-P495 is $7,399 on sale, with a list price of $9,249.

Again, the new machine brings more than extra memory. It gets the PRO 495, faster LPDDR5X, updated graphics, newer platform tuning and features built around multi-node deployment.

But the basic pattern is the same. A 128GB-class Ryzen AI Max+ 395 machine lives around the mid-$3,000 range, while the newest 192GB systems are clustering around $6,400 to $7,400 before we even look at the highest list prices.

That is why I do not think this is one vendor overcharging. The whole 192GB class is expensive.

The memory market is doing part of the damage

There is also terrible timing behind these products. LPDDR5X is expensive in 2026, and Framework has been unusually open about how much that is hurting its pricing.

Framework said on September 29 that LPDDR5X costs continue to increase, and it had to raise the price of its 128GB Desktop again as older, cheaper memory inventory ran out. The company also said it is opening only one initial batch of the 192GB model, with pricing and timing for later batches depending on memory cost and supply.

That is important context. The new systems did not arrive in a normal memory market.

AI infrastructure demand is pulling huge amounts of high-value memory capacity toward HBM and server products, while PC vendors are paying more for the large LPDDR pools they need for unified-memory AI systems. A 192GB mini PC is arriving at exactly the moment memory is one of the most painful parts to buy.

This does not make $6,799 feel cheap. It does explain why vendors are not treating the extra 64GB like a normal RAM upgrade.

The old 128GB machines suddenly look much better

This is the part I would pay attention to if I were shopping today. The arrival of 192GB systems may have made the 128GB Ryzen AI Max+ 395 machines more attractive, not less.

Framework’s current 128GB system costs $3,449. Minisforum’s 128GB MS-S1 MAX is $3,799 on sale. Both still give you a 40-CU Radeon 8060S, LPDDR5X-8000 and enough unified memory for a lot of serious local models.

Framework’s own published local-AI guide lists OpenAI gpt-oss-120b at around 38 tokens per second on its 128GB Max+ 395 configuration using LM Studio on Fedora 42 with MXFP4. That is not a toy workload.

If the models you actually use fit inside 128GB, spending another $3,000-plus to reach 192GB may buy very little day-to-day speed. You are mostly buying capacity you may or may not use.

That is not exciting advice, but it is probably the smart advice for a lot of buyers.

The 192GB machine is for workloads that cross a hard line

The easiest mistake here is asking whether 192GB is “worth it” in general. That question is too vague.

The better question is whether your workload crosses 128GB. If it does not, the older machine can be the better value even though the product name is older. It has the same basic GPU architecture, the same 40-CU count, and only slightly lower memory bandwidth.

If your workload needs 140GB, 155GB or 175GB of resident memory, the answer changes completely. Now the 192GB box lets you stay on one machine, while the 128GB box forces compromise.

That is why the price curve feels so brutal. Capacity upgrades near a hard limit are worth almost nothing to one buyer and thousands of dollars to another.

192GB does not automatically mean 192GB for the model

Another detail gets lost in the headline number. The operating system, runtime, KV cache, context, applications and background services all use the same unified-memory pool.

You cannot assume a 192GB machine gives a model a clean 192GB allocation. AMD’s platform can allocate very large portions of memory to graphics, and vendors are advertising graphics allocations well above 100GB, but some headroom still matters.

Long context can consume a lot of cache. Running two models at the same time changes the equation again.

That is exactly why 192GB can be useful even for a model whose quantized weights are well below 192GB. The extra room is not wasted if it lets you keep longer context or multiple agents resident.

But the same logic works in reverse. If your entire real workload sits happily under 100GB, buying 192GB because the number looks future-proof may be an expensive form of anxiety.

The “GPU barely changed” headline is not saying performance never improves

I want to be precise here because the title is deliberately sharp. The newer PRO 495 system can absolutely be faster than a Max+ 395 machine.

The GPU clock is higher. The memory is faster. CPU clocks move a little, and firmware, drivers and inference software keep improving too.

Framework itself says those changes should improve AI inference speed over the original Desktop. I believe that.

What I do not see in the specifications is a hardware change large enough to explain a near-doubling of complete-system price as a speed upgrade. The new price is mainly understandable when you treat 192GB capacity as the premium feature.

That is the real point.

Local AI is creating a very strange PC market

Normal PC buying used to have a fairly simple shape. Spend more money and you generally got more CPU cores, a larger GPU, more storage, or some mix of all three.

Local AI is breaking that pattern. Now a machine can cost nearly twice as much because it has 64GB more soldered memory, while the GPU underneath is still 40 compute units from the same architecture generation.

That feels wrong if you think like a gamer. It makes much more sense if you think like somebody buying a database server or a memory-heavy workstation, because the capacity itself is the feature.

Mini PCs are drifting into that world. The box is small, but the buying logic is starting to look like enterprise hardware.

This is why comparing only TOPS is becoming useless

The new PRO 495 platform is often marketed with 131 TOPS of total AI performance. The number looks large and tidy, which is why vendors like putting it on product pages.

For local LLM inference, I would rank it below memory capacity, memory bandwidth, runtime support and actual token-generation benchmarks. A 131-TOPS machine with 192GB can be more useful than a higher-TOPS device if the bigger model fits only on the 192GB machine.

At the same time, a much faster NVIDIA GPU with 32GB can crush it on a model that fits fully in VRAM. There is no one-number AI benchmark anymore.

That is annoying, but at least it makes hardware interesting again.

I would buy 128GB unless I already knew why I needed 192GB

That is where I land after looking at the prices. If I were buying a local-AI mini PC today and my normal models fit in 128GB, I would seriously look at the discounted 395 machines first.

A Framework at $3,449 or Minisforum at $3,799 leaves more than $3,000 in the budget compared with many 192GB options. That money can buy storage, networking, a second small machine, cloud credits for occasional giant-model jobs, or simply stay in your bank account.

If I already had a specific 150GB-class model, multi-model workflow or agent stack that needed the extra memory, I would buy 192GB and stop pretending the 128GB option was equivalent. The key is knowing which buyer you are before clicking checkout.

That distinction sounds boring. It may save you several thousand dollars.

The new mini PCs are bigger in memory, not twice as fast

The Ryzen AI Max+ PRO 495 is a useful update. It raises the memory ceiling to 192GB, moves LPDDR5X from 8000 to 8533MT/s, nudges the Radeon GPU from 2.9GHz to 3.0GHz, and gives vendors a platform for much larger single-box local-AI workloads.

What it does not do is double the GPU. The compute-unit count stays at 40, Framework says theoretical memory bandwidth rises only about 6.7%, and the architecture stays in the same family.

Yet complete machines can jump from around $3,500 for 128GB to around $6,800 or $7,000 for 192GB. That sounds ridiculous until your model needs 140GB, because then the extra 64GB becomes the only spec on the page that matters.

That, more than the new processor name, is what the 192GB mini-PC era is really selling.