If you are buying a mini PC to run AI privately in your own office, AMD’s Ryzen AI Max+ chips are the ones most small machines are built around. There are now two generations on sale: the 395 and the newer 495. The short answer is that the 495 is the same engine with a much bigger fuel tank. It is only slightly faster, but it can hold 192GB of memory against 128GB, and for local AI, memory decides which models you can run at all.
The differences at a glance
| What | Ryzen AI Max+ 395 | Ryzen AI Max+ 495 |
|---|---|---|
| Cores and threads | 16 cores, 32 threads (Zen 5) | 16 cores, 32 threads (Zen 5) |
| Clock speed | 3.0GHz base, up to 5.1GHz | 3.1GHz base, up to 5.2GHz |
| Graphics | Radeon 8060S, 40 compute units | Radeon 8065S, 40 compute units |
| NPU | Up to 50 TOPS (126 TOPS across the whole platform) | Up to 55 TOPS (131 TOPS across the whole platform) |
| Maximum memory | 128GB LPDDR5x-8000 | 192GB LPDDR5x-8533 |
| Memory a model can use as graphics memory | Up to 96GB (vendor figure) | Up to 160GB (vendor figure) |
| Power | 45W to 120W configurable | 45W to 120W configurable |
These are AMD’s published figures as reported in ACEMAGIC’s buying guide, plus two vendor figures that we have marked as such. We have not benchmarked the two chips side by side ourselves.
What actually changed
Look down the table and most rows barely move. Both chips have 16 cores, the same graphics design with 40 compute units and the same power range. The 495 gains 100MHz of clock speed and a few points of NPU performance. You would struggle to notice either at a desk.
The row that matters is memory. These chips use unified memory, which means the processor and the graphics share one pool, and a large language model has to fit inside that pool to run at a useful speed. Moving from 128GB to 192GB, with faster memory as well, is the whole point of the 495.
Why memory matters more than speed for private AI
A model that does not fit in memory either will not load or runs painfully slowly. More memory gives you three practical things:
- Larger models. Bigger models generally reason better and make fewer mistakes on long, multi-step work.
- More than one model at once. Agent teams often pair a large model for thinking with a small, quick one for routine steps and another for searching your documents.
- More context. Long documents, long conversations and several people using the machine together all consume memory on top of the model itself.
Vendors pitch 395 machines at models of around 70 billion parameters and 495 machines at around 120 billion, with larger “mixture of experts” models possible on both. Treat those as capability targets, not measured speeds. The same vendor guide says plainly that it gives no tokens-per-second results, and you should be wary of anyone who quotes a single speed figure without naming the model and settings.
Which one should a business buy?
Choose a 395 machine if you are proving the idea in one department, a handful of people will use it, and a mid-sized model is enough. It is the cheapest way to get real private AI on a desk, and the machine stays useful as a second node later.
Choose a 495 machine if you want agent teams working for a whole workgroup, you expect to run a large model alongside smaller ones, or you want headroom so the hardware is not the limit a year from now. This is the class we specify for a workgroup.
Look beyond both if you need the NVIDIA software stack, or many people will use the system heavily at the same time. That is where NVIDIA’s DGX Spark class and larger systems come in.
The machines you can buy
On the 395 side there are compact boxes from GMKtec, Minisforum and Beelink, and HP’s Z2 Mini G1a for buyers who want a workstation warranty and manufacturer security. On the 495 side we list three: the Minisforum MS-S1 Max P495, the GMKtec EVO-X5 Pro and, new this week, the Beelink GTR9 Pro 495.
The Beelink is worth a note. It is the only one of the three offered with 128GB as well as 192GB, it has dual 10GbE networking and a three-year manufacturer warranty. It is currently a pre-order that ships from China in around five weeks, with UK import duty and VAT to add, so we quote a landed price instead of showing one.
Specifications and current prices for every model are on our AI hardware page.
Hardware is the small part
The box is the base platform. What makes private AI useful is what goes on it: the models, the agents, the connections to your systems, the access controls and the training for your people. Two things we have learned from building these systems:
- Check the AI is really using the graphics. On these AMD chips we found that running the model server inside a container on Windows quietly fell back to the processor and ran several times slower. Running it directly on the machine fixed it. Nothing warned us; we only saw it by measuring.
- Bigger is not always better. For agents that have to call tools reliably, we have seen a smaller model outperform a larger one. Size the model to the job, then size the memory to the models.
That is why we supply hardware as part of a build, not as a box on its own: we specify it, install it, connect it and train your team to use it.
How Total Group can help
Total Group builds private AI that runs on hardware you own, so your data stays in your business and your costs stay predictable. We start with an AI Readiness Assessment, because AI should only see what it is supposed to see.
Talk to us about the right AI hardware for your business or call 01727 881 224.
Sources
- Ryzen AI Max+ 395 vs PRO 495 buying guide (acemagic.uk), for AMD’s published specifications and the vendor memory figures
- Beelink GTR9 Pro, Ryzen AI Max+ 495 (bee-link.com), for the Beelink configuration, warranty and shipping details
Frequently asked questions
What is the difference between the Ryzen AI Max+ 395 and 495?
They share the same 16-core design, graphics layout and power range. The 495 is slightly faster, but the real change is memory: it supports up to 192GB of faster unified memory against 128GB on the 395, which lets it run larger AI models or several models at once.
Is the Ryzen AI Max+ 495 worth the extra cost for local AI?
It is if you need the memory. For a single department trying private AI with a mid-sized model, a 395 machine is usually enough. For agent teams serving a workgroup, or for running a large model alongside smaller ones, the 192GB of a 495 machine gives useful headroom.
How much memory do I need to run AI models locally?
The model has to fit in memory with room left for the conversation and documents it is working on. Mid-sized models run comfortably in 64GB to 128GB of unified memory. Larger models, or several models running together for agent work, are where 192GB starts to matter.
Can a mini PC really run private AI for a business?
Yes, for a department or workgroup. A mini PC with a Ryzen AI Max+ chip can run capable open models entirely on your premises, so data does not leave the business. Heavier use by many people at once calls for NVIDIA DGX class hardware or a GPU server.