Ryzen AI Max+ PRO 495 192GB (2026): Wait or Buy 128GB?

Bottom Line

• Gorgon Halo is a capacity upgrade, not a speed upgrade. Memory goes from 128GB to 192GB. Bandwidth goes from 256 to 273 GB/s, +6.6%. The GPU is the same 40-unit RDNA 3.5 part.
• No 192GB box has a price as of 2026-09-16. Seven vendors have shown one. Minisforum’s store page reads ”$????”.
• The 128GB boxes are not cheap any more. They now cost $3,450 to $4,350, and they are mostly in stock.
• Wait only if your target model is 100GB to 155GB. For everything smaller, you would pay more for the same tokens per second.
• “495” does not mean 192GB. Lenovo’s 495 machine stops at 128GB. Check the exact configuration.

What Changed From the Max+ 395

Specifications are from AMD’s product pages, read 2026-09-16. AMD does not publish a GB/s figure. The bandwidth and GPU-memory rows come from vendor pages and press coverage.

Token generation on these machines is limited by memory bandwidth. A 6.6% rise in bandwidth gives, at best, a similar rise in tokens per second. No independent tok/s test of a 495 machine exists yet. Notebookcheck has Geekbench scores only. We will not guess a number.

All Three Chips Are PRO, and All Three Reach 192GB

AMD lists three parts: the Max+ PRO 495 (16 cores, 40 GPU units), the Max PRO 490 (12 cores, 32 GPU units) and the Max PRO 485 (8 cores, 32 GPU units). All three list 192GB as the maximum. Tom’s Hardware reports that AMD has confirmed only the PRO series, and that a consumer version is undecided.

For local LLM work, buy the 495 if you have the choice. The 490 and 485 have a smaller GPU, and prompt processing depends on GPU compute.

What 160GB Runs That 96GB Does Not

We read these GGUF sizes from the Hugging Face API on 2026-09-16. Sizes are decimal GB. Leave room for the KV cache on top.

That is the honest list: a small set of 130-155GB MoE models at 4-bit. It is a real tier. It is also a narrow one.

Two marketing claims do not survive the file sizes:

• “Fits DeepSeek-V4-Flash at Q8.” The Unsloth Q8-class file is 161.9GB. The GPU can use 160GB. It does not fit with any context.
• “Handles 300B+ models.” Only at 3-bit. Qwen3.5-397B-A17B needs UD-IQ3_XXS (140.3GB) to fit. Minisforum’s own page shows that model on two clustered units, at a vendor-stated 16 tok/s with no quant named.

For GLM-5.3-Flash there is a second problem: stock llama.cpp could not load it as of mid-September. See Can I run GLM-5.3-Flash locally?.

Who Has Shown a 192GB Box

Status as of 2026-09-16. Every row was read on the vendor page or in dated press coverage.

About the “€7,000” figure. A VideoCardz headline gave that price for the Minisforum box. The article body says Minisforum has not announced pricing, and that a teaser image only appears to suggest it. Treat it as unconfirmed.

About the “$2,300-2,700” guesses. They are below what a 128GB box costs today. A machine with 50% more soldered LPDDR5x will not cost less than the smaller one. Framework has said to expect “a pretty substantial jump” over its 128GB price, as quoted by Notebookcheck.

AMD’s own $3,999.99 “Ryzen AI Halo” box is not Gorgon Halo. It uses the Max+ 395 with 128GB, and it sells at Micro Center for in-store pickup only. A 495 version is “coming soon” with no date.

What a 128GB Box Costs Now

We read these prices on each vendor’s own store on 2026-09-16.

In mid-August, GMKtec’s store listed the EVO-X2 at $1,999.99 and showed every variant as sold out. In mid-September it is in stock at $3,499.99. The cheap price was never a price you could pay. GMKtec’s store now carries a notice that another increase is coming.

This matters for the wait-or-buy decision. The 2026 DRAM shortage moves these machines up, not down. The memory is soldered and it is most of the cost. For the full comparison, see which Strix Halo mini PC to buy.

Wait or Buy

Buy a 128GB box now if:

1. Your models fit in about 96GB of GPU memory. That covers gpt-oss 120B and the 4-bit files under 100GB.
2. You want a price you can see. No 192GB price exists to compare against.
3. You need the machine this quarter.

Check the GMKtec EVO-X2 128GB on Amazon

Check the Minisforum MS-S1 MAX 128GB on Amazon

Wait for a 192GB box if:

1. You have a named model between 100GB and 155GB that you need at 4-bit, such as DeepSeek-V4-Flash or MiniMax-M2.7.
2. You accept that it will run at Strix Halo speed. The extra memory makes a larger model load. It does not make it fast.
3. You can wait for a price, and then for stock. In this category the two have not arrived together.

Skip both if your budget is above $4,000 and you want CUDA. The DGX Spark Founders Edition lists at $4,699 as of September 2026, with 128GB of unified memory and NVIDIA’s software stack. A Beelink GTR9 Pro at $4,349 is $350 below it. See DGX Spark vs Strix Halo.

Check the NVIDIA DGX Spark on Amazon

Sources

• AMD product pages: Ryzen AI Max+ PRO 495, Max PRO 490, Max PRO 485, and Ryzen AI Max+ 395, read 2026-09-16
• Tom’s Hardware: “AMD Ryzen AI Max 400 ‘Gorgon Halo’ packs up to 192GB of unified memory” (published 2026-05-21)
• Framework Desktop 192GB page (192GB, 273 GB/s, “coming soon”), read 2026-09-16
• Minisforum store: MS-S1 MAX-P495 page (price placeholder, clustered Qwen3.5-397B claim) and MS-S1 MAX page ($3,799), read 2026-09-16
• VideoCardz: Minisforum P495 (2026-09-05), Framework 192GB (+6.6% bandwidth), AMD Ryzen AI Halo $3,999.99, Lenovo ThinkCentre X Ultra, CHUWI, AOOSTAR, COLORFIRE coverage
• Notebookcheck: Framework 192GB Desktop (2026-07-22) and Ryzen AI Max+ PRO 495 benchmark page (Geekbench only)
• TechRadar (2026-09-07): no final retail pricing from Framework, GMKtec or Acemagic
• gmktec.com, bee-link.com and frame.work store pages, read 2026-09-16
• Hugging Face API file trees for the Unsloth GGUF repos named in the table, read 2026-09-16
• NVIDIA DGX Spark $4,699 list: NVIDIA developer forum price notice, re-verified 2026-09-13

See Also

• Which Strix Halo mini PC to buy
• Best models for the AMD Ryzen AI Max+ 395
• DGX Spark vs Strix Halo for local LLMs
• DGX Spark price: every GB10 box
• Best local LLMs for 128GB RAM
• DeepSeek V4-Flash vs gpt-oss-120b
• Should you buy RAM now or wait out the 2026 shortage?