A compact desktop computer sits beside a monitor illustrating a 300-billion-parameter AI model and shared memory across processors.
The "Ryzen AI Max Plus Pro 495" mini PC in the latest ETA Prime video coverage is AMD's refreshed Ryzen AI Halo Developer Platform. AMD introduced it on October 12, 2026, powered by the Ryzen AI Max+ PRO 495 processor. The headline spec is 192 GB of unified memory. A few details in the viral coverage need correcting, and the buying picture is more specific than "a mini PC with lots of RAM."

A compact desktop computer sits beside a monitor illustrating a 300-billion-parameter AI model and shared memory across processors. What AMD actually launched​

AMD calls this a developer platform, not just a processor. AMD says the up-to-192 GB memory pool, support for 300-billion-parameter and larger models, and a curated software experience are meant to let developers and creators build, experiment and run more AI locally.

The core hardware, per AMD's launch material:

  • Up to 16 "Zen 5" CPU cores, 40 RDNA 3.5 GPU compute units, an NPU of up to 55 TOPS, and up to 192 GB of LPDDR5x unified memory.
  • HotHardware reports 16 cores and 32 threads, from a 3.1 GHz base clock to a 5.2 GHz maximum boost.
  • Memory is LPDDR5X-8533, with 160 GB allocated as video memory.
  • Micro Center's listing shows a 2 TB M.2 SSD (SED), 10GbE, Wi-Fi 7 and Bluetooth 5.4.

Corrections to the circulating story​

The GPU name. The Geeky Gadgets piece calls the graphics "Radeon 865S." Micro Center's listing names it AMD Radeon 8065S. Use that designation.

The operating system. The coverage says the box ships with Linux and that Windows users must install it themselves. That is only partly right:

  • The Linux version is priced at $7,299.99, with a Windows version to follow.
  • AMD says Windows-based Ryzen AI Halo platforms with the same processor and Microsoft's Windows Project Zenith are coming soon.

For Windows shops, the practical answer is to wait for the Windows SKU. I haven't seen a source that confirms its pricing or date. I also haven't verified how a Linux-shipped unit behaves if you wipe it and load Windows.

The memory. "192 GB of RAM" does not mean 192 GB of GPU memory. The CPU, GPU and NPU share one memory pool, and up to 160 GB of it can be allocated to the GPU as VRAM. The rest stays with the system.

The 300B-parameter claim​

AMD's headline is local large language models. AMD says the platform can support 300-billion-parameter and larger language models at 4-bit quantization. It also says developers can run multiple large agents at once with discrete KV caches.

The arithmetic is plausible. Tech Insider notes that a quantized model's weights take roughly half a byte per parameter at 4 bits. That puts a 300B model's weights at around 150 GB, which fits inside a 160 GB allocation. It leaves little headroom for context, KV cache and the OS.

This is a vendor capacity claim. It says nothing about tokens per second, long-context behavior or software compatibility. It also does not show that the box can replace the cloud, or that it is good at training. The ETA Prime coverage leans on "training and inference" language, but AMD's stated pitch is running and experimenting locally. Memory bandwidth is the usual limit on how fast large models generate tokens on unified-memory systems. That is general industry knowledge, not something I measured here. Expect a big model to load but run slowly compared with a datacenter GPU.

Claims I couldn't verify​

The Geeky Gadgets piece relays several figures I could not confirm against AMD or retail sources:

  • 8.1% single-core and 10.9% multi-core gains over the Max+ 395.
  • 10.2% synthetic and 11.9% real-world GPU gains.
  • 1440p high-settings gaming in named titles.
  • Parity with an RTX 4060 laptop GPU in some workloads.
  • A 240W power supply and "four USB-C ports."

These come from one YouTuber's testing, relayed through an affiliate-supported site. Treat them as one reviewer's results, not a confirmed spec sheet. Gaming is also a side note for a $7,300 developer box.

Price and who it's for​

Back2Gaming says it can be ordered now from Micro Center. At the Linux price of $7,299.99, this is a niche tool.

  • Good fit: teams that need big local models for data-privacy reasons, developers prototyping multi-agent workflows, and labs testing 300B-class open-weight models without renting GPU time.
  • Poor fit: anyone who wants a gaming mini PC, or anyone who needs a Windows-first experience today.

The tradeoff against a discrete-GPU rig is capacity versus speed. A multi-GPU workstation will usually generate tokens faster, but at this size it is typically costlier, louder and more power-hungry. That is a general industry comparison, not a measured one.

Bottom line​

The Ryzen AI Halo refresh is a real product with a concrete spec change. The memory ceiling goes up to 192 GB, with up to 160 GB usable as GPU memory, and AMD says that is enough for 300B-class models at 4-bit quantization. Windows buyers should wait for the Project Zenith version. Everyone should hold the performance percentages loosely until independent reviews arrive.

 

References

  1. AMD Ryzen AI Max Plus Pro 495 Packs 192GB RAM for Local AI - Geeky Gadgets Geeky Gadgets 2026-10-24T14:00:00+00:00
  2. AMD Ryzen AI Halo Now Packs Ryzen AI Max+ Pro 495 And 192GB Memory hothardware.com
  3. AMD's Next Ryzen AI Halo Packs Ryzen AI Max+ PRO 495 and 192GB of Unified Memory - Back2Gaming back2gaming.com