Ryzen AI Max+ 495 vs Max+ 395: Are 192GB Mini PCs Worth Double?

For most people, the 192GB Ryzen AI Max+ PRO 495 is the wrong buy. It costs about twice as much as a 128GB Ryzen AI Max+ 395 box. It uses the same Zen 5 silicon. It generates text only a few percent faster. What it changes is how big a model fits, and that matters only inside a narrow size window.

This page compares the Ryzen AI Max+ 495 vs 395 on what decides the purchase. That means the spec differences, which models fit in 96GB, 120GB and 160GB of GPU memory, and bandwidth math for expected tokens per second. It also covers what the 495 boxes cost and when they ship. The five 395 machines are reviewed in our Ryzen AI Max+ 395 mini PC comparison. The best mini PC for local AI guide sets both chips against the DGX Spark and the Mac Studio. Prices were verified on Amazon US and the maker stores on 2 October 2026. Speed figures are cited or estimated from bandwidth, not measured by us, because no independent 495 benchmark had been published by then.

The verdict and where to buy

Buy a 395 unless your model needs between about 120GB and 180GB of GPU memory on Linux (96GB to 160GB if you run Windows). That band holds Qwen3-235B and MiniMax-M2 at Q4, plus Qwen3.5-397B at 3-bit. Below it, a 395 runs the same models at nearly the same speed for half the money. Above about 160GB on Windows, or about 180GB on Linux, neither chip is enough on its own.

No 495 machine is sold on Amazon US. The only 495 listings there are from NIMO. All three show as currently unavailable. They also quote 9600MHz memory and a 5.5 GHz boost, which the AMD spec sheet does not support. Skip them. Check the live price before ordering anything above. The EVO-X5 Pro has already dropped from its $6,799 launch price, and memory-heavy machines are repricing often.

What AMD changed in the 495

Very little. The Register calls the 495 essentially a factory-overclocked Strix Halo, and the two AMD spec pages agree. Same 16 Zen 5 cores, same 40 GPU compute units, same 256-bit memory bus, same 55W default TDP, same FP11 package.

The AMD product page does not list bandwidth. The 273 GB/s figure is 8533 MT/s across a 32-byte bus, and The Register, CHUWI and Hardware Busters all quote it. The PRO badge adds the AMD business manageability stack (DASH, Memory Guard, Secured-core). Home users will not use it, and they lose Precision Boost Overdrive and Curve Optimizer in exchange.

What 192GB buys you

The extra 64GB is the only reason to pay for a 495. How much of it the GPU can use depends on the OS.

On Windows the split is fixed in firmware: 96GB of GPU memory on a 128GB 395, and 160GB on a 192GB 495. Linux is more generous. A cluster guide from AMD sets amdgpu.gttsize=120000 and a matching ttm.pages_limit to give each 128GB 395 about 120GB of GPU memory. The Register reports that Linux can hand nearly all 192GB of a 495 to the GPU. That 120GB Linux figure matters. Several models that “need” a 495 on Windows fit a 395 on Linux at one quant step lower.

Sizes below are the full GGUF downloads from Hugging Face, all shards summed. “Tight” means the weights fit but less than 10GB is left for the KV cache, so context length suffers. The GPU memory caps are nominal binary figures, so the table errs on the conservative side.

The rows from MiniMax-M2 Q4 to GLM-4.6 Q3 are where the 495 earns its price. Those models run at Q4 or 3-bit with real context room on a 495 (GLM-4.6 Q3 is tight on Windows). They don’t load at all on a single 395. Everything above that block fits a 395 already, at least on Linux. Qwen3-Coder-480B at Q2 only squeezes onto a 495 under Linux, and anything bigger needs two machines whichever chip you pick.

Ignore claims that a 495 runs DeepSeek V4-Flash at Q8. Digital Trends reported that line from the Framework launch. The model has 284B parameters, so Q8 weights would be around 285GB. Its native checkpoint is a 159.6GB FP4 and FP8 mix, which only just fits the 160GB Windows allocation. Linux leaves room for context. Our VRAM sizing guide covers the per-quant math if your model is not listed.

How much faster the 495 generates

Barely. Token generation on these chips is limited by memory bandwidth, and the 495 adds 6.6 percent. The clock bumps do not help a workload that waits on DRAM.

There is no measured 495 number to quote yet. StorageReview held benchmarks of its HP ZBook Ultra G3a preview unit for shipping hardware. The only owner report we found, a Lemonade GitHub issue, confirms llama.cpp runs but gives no speeds. So the table below is arithmetic. Take the bytes each token reads (active parameters times the bytes per parameter of that quant) and divide them into peak bandwidth to get a ceiling. Then scale the ceiling by 58 percent. That is how close a 395 gets to its own ceiling on gpt-oss-120b, at the cited 52 to 56 tok/s.

Read these as rough short-prompt estimates. The 58% efficiency comes from one MXFP4 model, so other quant formats may land lower. They ignore KV cache reads, which grow with context, so long conversations land lower on both chips. Prompt processing is compute-bound, and with identical compute units it should be close to even.

The table hides a trade. On Linux, a 395 runs Qwen3-235B at Q3 slightly faster than a 495 runs it at Q4, because the smaller quant reads fewer bytes per token. What the 495 buys is quality and context room at the same model size, not speed. If 3-bit output is good enough for your work, the 395 already does the job. The llama.cpp setup guide covers the Vulkan and ROCm builds that produce the cited 395 numbers.

Two 395s instead of one 495

Two Bosgame M5 boxes cost about $6,648 on Amazon. That is roughly what one ACEMAGIC F9A costs. You get 256GB of RAM, about 240GB of it usable by the GPUs on Linux. The RPC backend in llama.cpp splits a model across machines over Ethernet. AMD published a guide doing exactly this with four 128GB Framework Desktops on 5GbE. They ran the trillion-parameter Kimi K2.5 at 9.45 tok/s with flash attention on.

The catch is the network. Every token crosses the link between nodes, so a split model generates slower than it would on one box with enough memory. The Bosgame and EVO-X2 each have a single 2.5GbE port. A pair of Minisforum MS-S1 Max boxes with dual 10GbE costs about $7,598. That is $200 more than one MS-S1 MAX-P495. It buys 64GB more memory plus a second machine to keep running if one fails. Two boxes also mean two power supplies, two OS installs and an RPC server to babysit.

Pick the single 495 when the model fits in 160GB (about 180GB on Linux) and you want one quiet machine. Pick two 395s when you need more than 160GB on Windows or roughly 180GB on Linux, or when you would rather split smaller models across two machines than run one big one.

Linux and ROCm support

The 495 GPU is the same gfx1151 target as the 395, so Strix Halo tooling carries over unchanged. The AMD ROCm compatibility matrix lists the Ryzen AI Max+ PRO 495 (Radeon 8065S) as gfx1151 and supports Ubuntu 26.04 and 24.04 on these APUs. GTT tuning carries over, ROCm builds such as those from the Lemonade project target the same gfx1151 ID, and Vulkan builds of llama.cpp need no change.

The same kernel rule applies to both chips. Run kernel 6.18.4 or later, or the Ubuntu linux-oem-24.04 kernel, for the CWSR fix that stops ROCm crashing on gfx1151. Don’t install amdgpu-dkms on these APUs. One naming bug hits only the 495. The Windows ROCm backend in Lemonade rejects “8065S” as an unsupported GPU (issue 3662). A native llama.cpp ROCm build works. The pre-built Framework 495 ships with Fedora and the HP ZBook Ultra G3a offers Ubuntu 26.04, so Linux is a first-class option on the early machines.

Before you order a 495

Every 495 box is either on pre-order or shipping in batches. GMKtec sold out its first 100 EVO-X5 Pro units, and new orders ship from 9 October. Minisforum quotes mid October. ACEMAGIC is a late October pre-order. Framework priced its 495 at $6,799 DIY and $7,449 pre-built, per Phoronix. TechPowerUp reports the first batch sold out, and it reportedly ships in November. Two more matter less to US buyers. The 192GB Bosgame M5 Max is due in early 2027. The Beelink GTR9 Pro 495 is sold only in China, with 128GB and the same 96GB GPU cap as a 395.

The first independent tokens-per-second numbers should arrive with shipping hardware this month. If your model sits in the 120GB to 180GB Linux band (96GB to 160GB on Windows), order now. You are paying for the memory, not the benchmark. If it does not, a 128GB 395 from the table at the top ships from Amazon stock today for about half the price. Our 395 comparison covers how those boxes differ in ports, cooling and noise.