The Ryzen AI Max+ PRO 495 is the most expensive way to buy the chip that is already in dozens of cheaper mini PCs. As of September 29, 2026, the cheapest PRO 495 mini PC you can order costs $6,499. For $6,198 you can buy two Ryzen AI Max+ 395 mini PCs with 128GB each, which is 256GB of memory, 4TB of storage and two working computers. And a 256GB Mac Studio costs $9,499, about $3,000 more than the PRO 495, but moves data through memory 4.4 times faster.
The PRO 495 sits in the worst spot between those two options. It is priced like a premium machine and performs like the budget one. Below is the math, the parts of the argument that do not favor us, and the short list of buyers who should still get one.
What a PRO 495 mini PC costs today
These are the prices on each brand’s own store, checked on September 29, 2026. All three ship with 192GB of memory.
| PRO 495 mini PC | Memory / SSD | Price | Status |
|---|---|---|---|
| ACEMAGIC F9A PRO 495 | 192GB / 2TB | $6,499 ($6,999 “regular”) | Orderable |
| GMKtec EVO-X5 Pro | 192GB / 2TB | $6,599 ($6,799 “regular”) | Orderable |
| Minisforum MS-S1 Max-P495 | 192GB / 2TB | $7,399 ($9,249 “regular”) | Orderable, ships mid October |
The GMKtec figure is already higher than its launch. On September 28 the EVO-X5 Pro opened at a $6,399 “Super Early Bird” price; a day later the store lists $6,599. ACEMAGIC being the cheapest comes with history: the brand shipped mini PCs with preinstalled malware in 2024, which matters more on a $6,499 order than on a $300 one. The only cheaper PRO 495 desktop, the Lenovo ThinkCentre X Ultra at a reported $3,699, tops out at 128GB, so it gives up the one thing the PRO 495 adds.
What the same money buys instead
Two Ryzen AI Max+ 395 mini PCs. The Bosgame M5 with 128GB and 2TB costs $3,099 on Bosgame’s store. Two of them cost $6,198, $301 less than the cheapest PRO 495. ACEMAGIC’s own M1A PRO+ with the 395, 128GB and 2TB also sells for $3,099 on the same store that asks $6,499 for the F9A PRO 495. Two GMKtec EVO-X2 units at $3,649.99 each come to $7,300, which is still cheaper than one Minisforum P495.
A Mac Studio with 256GB. Apple’s current Mac Studio uses the M5 Ultra. The 256GB configuration with a 1TB SSD costs $9,499; with 2TB, $9,999. Apple rates its memory bandwidth at 1.2TB/s.
| Option | Total memory | Memory bandwidth | Price | Price per GB of memory |
|---|---|---|---|---|
| One Ryzen AI Max+ PRO 495 (ACEMAGIC F9A) | 192GB | ~273GB/s | $6,499 | ~$34 |
| Two Ryzen AI Max+ 395 (Bosgame M5) | 2 × 128GB | ~256GB/s each | $6,198 | ~$24 |
| Mac Studio M5 Ultra, 256GB / 1TB | 256GB | 1.2TB/s | $9,499 | ~$37 |
Look at the last column next to the bandwidth column. Per gigabyte, the PRO 495 is priced almost like the Mac Studio, $34 against $37. Per unit of speed it is nowhere close: the Mac delivers about 4.4 times the memory bandwidth for 46% more money. The two 395s give you a third more memory than the PRO 495 for less money.
Why memory bandwidth decides the argument
When a local language model writes an answer, each new token requires reading the model’s active weights from memory. The practical ceiling on generation speed is therefore roughly memory bandwidth divided by the gigabytes read per token. That is why a dense 70B model at 4-bit, about 40GB read per token, is capped at roughly 6 tokens per second on a 395 (256 ÷ 40), while OpenAI’s gpt-oss-120b, a mixture-of-experts model that reads only about 5 billion active parameters per token, ran at 33 tokens per second in Jeff Geerling’s test on the same chip.
The PRO 495 raises memory capacity by half (128GB to 192GB) but raises bandwidth by about 6.6% (LPDDR5X-8000 to 8533, roughly 256GB/s to 273GB/s). That creates a mismatch. The extra 64GB lets you load bigger models, but it does not help the box read them faster.
A simple example, calculated from the formula above and not measured: a dense model that fills 150GB of the PRO 495’s memory has a ceiling of about 1.8 tokens per second (273 ÷ 150). You can load it. You will not want to use it. The Mac Studio’s ceiling for the same model is about 8 tokens per second.
The PRO 495’s extra memory is useful only for mixture-of-experts models that are large on disk but read a small slice per token, and for long contexts that need room for the KV cache. That is a real use case, and we come back to it at the end. It is also exactly the use case that two 395s handle well.
The chip itself: a 395 with more memory support
AMD did not build a new chip. The PRO 495 uses the same Zen 5 CPU, RDNA 3.5 GPU and XDNA 2 NPU as the 395, with small clock increases, as our Ryzen AI Max 400 “Gorgon Halo” explainer details:
- CPU: up to 5.2GHz instead of 5.1GHz. PassMark’s CPU Mark puts it about 4% ahead in multi-thread and about 3% ahead single-thread.
- GPU: Radeon 8065S at up to 3.0GHz instead of the 8060S at 2.9GHz, with the same 40 compute units.
- NPU: 55 TOPS instead of 50.
- Memory: up to 192GB of LPDDR5X-8533, with up to 160GB assignable to the GPU, against 128GB and roughly 96GB (Windows) to 120GB (Linux) on the 395.
If you do not run AI models, none of this is worth $3,000. For gaming, video editing or software builds, a 4% CPU gain and a 100MHz GPU bump are within the difference between two 395 boxes with different cooling.
What 128GB already runs
We run a Minisforum MS-S1 Max with the Ryzen AI Max+ 395 and 128GB in our own lab. With llama.cpp on Vulkan, Qwen3.5-122B-A10B at Q4_K_XL (77GB of weights) generates 21 to 22 tokens per second with around 275 tokens per second of prompt processing, and leaves room for a 32K context. A 120B-class mixture-of-experts model is the sweet spot for this hardware, and it does not need a PRO 495.
Add gpt-oss-120b at 33 tokens per second from the section above: the popular 120B-class open models already fit comfortably in 128GB.
The case for two 395s, and its costs
The strongest argument for two 395s comes from AMD. Its own playbook, “Clustering Two Ryzen AI Halos with RPC”, runs GLM 4.7, a 358-billion-parameter model at about 205GB, split across two 128GB Strix Halo machines with llama.cpp’s RPC mode over 10GbE. That model does not fit in a single PRO 495, which can give the GPU at most 160GB. AMD’s longer guide scales the same method to four machines for a one-trillion-parameter model.
Two machines also give you options a single box does not:
- Run two models at once, for example a coding model on one and an image or speech model on the other, each at full speed.
- Keep working when one fails. For a $6,000 purchase, having a second machine is useful insurance, especially with small brands whose support is inconsistent.
- Buy in two steps. Start with one 395 at about $3,100, add the second only when your models outgrow 128GB.
Now the costs, because clustering is not free:
- Speed drops when a model is split. Jeff Geerling’s gpt-oss-120b fell from 33 to about 24 tokens per second when spread across his cluster, and Llama 3.1 405B across four boards ran at 0.7 tokens per second. LTT Labs titled its two-machine test “World’s First(?) Underwhelming AMD Ryzen AI Halo Cluster”. Split a model across machines only when it does not fit in one.
- More setup. llama.cpp RPC works, but it is manual. You need a fast link (10GbE, 5GbE or a USB4 network cable), two operating systems to maintain, and the same llama.cpp build on both.
- Two power supplies, twice the desk space, and roughly twice the idle power draw.
So the fair summary is: two 395s beat one PRO 495 on price, total memory and flexibility, and lose on simplicity. For a model between 120GB and 160GB that must live on one machine, the PRO 495 is simpler. For everything smaller, one 395 is enough. For everything bigger, the PRO 495 cannot do it at all.
The case for the Mac Studio, and its costs
If you are going to spend $6,500 on local AI and bandwidth is what limits you, the Mac Studio M5 Ultra is the machine the PRO 495 cannot answer. At 1.2TB/s it reads memory about 4.4 times faster, and Tom’s Hardware’s review found its token generation almost four times faster than NVIDIA’s DGX Spark, a machine in the same bandwidth class as Strix Halo.
The honest limits:
- It costs $3,000 more than the cheapest PRO 495. That is not “the same money”. It is the next step up.
- macOS only. No CUDA, no ROCm, no Linux as the main OS. The MLX ecosystem and llama.cpp’s Metal backend cover most local LLM work, but not every tool.
- Memory prices went up. A year ago the gap was smaller: the M3 Ultra Mac Studio with 256GB cost $5,599 at its March 2025 launch, and Apple raised the 256GB upgrade by $400 in March 2026. The M5 Ultra starts at $5,499 with 96GB, and going to 256GB adds $4,000.
- No upgrades and Apple’s repair pricing, same as the soldered memory in every Strix Halo box.
Even at $9,499, the Mac Studio makes the PRO 495’s price look worse, not better. The PRO 495 asks Mac-level money per gigabyte while delivering the same bandwidth as a $3,099 Bosgame.
Who should still buy a Ryzen AI Max+ PRO 495
There is a real, narrow group:
- You need 130 to 160GB of GPU-addressable memory in one machine, for example a large mixture-of-experts model at Q4 plus a long context, and you will not run a cluster.
- Your company buys AMD PRO platforms for their fleet-management features and wants the largest configuration in a single, supported box. In that case, compare the Lenovo ThinkCentre X Ultra’s support terms with the smaller brands’ before anything else.
- You cannot use macOS, you need more than 128GB, and your desk or rack has room for one box, not two.
If none of those describe you, the PRO 495 means paying about $3,400 extra for 64GB of memory that runs at the same speed as the memory you already get with a 395.
Our verdict
We do not recommend buying a Ryzen AI Max+ PRO 495 mini PC at current prices. For most buyers, one Ryzen AI Max+ 395 with 128GB, such as the Bosgame M5 at $3,099, runs the models that matter today at half the price. If you really need more than 128GB, two 395s cost less than one PRO 495 and hold 256GB. If what you want is speed, not capacity, save for a Mac Studio.
Prices for the PRO 495 will likely fall once the launch rush passes and more brands, including Framework, publish theirs. At around $4,650, a 192GB PRO 495 would match the Bosgame M5’s roughly $24 per gigabyte, and the simplicity of one box would start to justify it. At $6,499 it does not.
Sources and caveats
PRO 495 prices come from the ACEMAGIC, GMKtec and Minisforum US stores and their product data, checked on September 29, 2026. Bosgame M5 and ACEMAGIC M1A PRO+ prices come from their official stores the same day, and the EVO-X2 price from GMKtec’s store. Mac Studio prices are from Apple’s US configuration pages on the same date; bandwidth and launch details from Apple’s August 2026 announcement. The March 2026 upgrade price change is MacDailyNews’s report. The 395 cluster setup is AMD’s own playbook; cluster performance figures are Jeff Geerling’s and LTT Labs’. PassMark figures come from its comparison page. The 122B-model numbers are our own measurement on a Minisforum MS-S1 Max (395, 128GB) with llama.cpp Vulkan. The bandwidth ceilings are arithmetic, not benchmarks. Store prices can change without notice.