Gorgon Halo Mini PCs: the $6,500 Price War Is Already On
AMD's Gorgon Halo platform — the Ryzen AI Max+ 400 series — went from one overpriced box to a four-way price war in a single week. If you want 192GB of unified memory for local AI, the entry price is now $6,499 and falling.
GMKtec's Evo-X5 Pro — the first Gorgon Halo box to go on sale. Photo: GMKtec.
What Gorgon Halo actually is
The Ryzen AI Max+ Pro 495: 16 Zen 5 cores and 32 threads boosting to 5.2GHz, Radeon 8065S integrated graphics (40 compute units, RDNA 3.5), and an XDNA 2 NPU rated at 55 TOPS — 131 TOPS for the platform overall. The headline feature is memory: 192GB of LPDDR5X-8533 unified memory, up to 160GB of which can be assigned as VRAM, at around 273GB/s of bandwidth.
The honest part: Gorgon Halo is a refresh of last year's Strix Halo on the same Zen 5 silicon, not a new architecture. Notebookcheck reports only minor performance gains over the older Ryzen AI Max+ 395. The real change is the memory ceiling doubling from 96GB to 192GB — which is exactly what the local-LLM crowd cares about.
The four boxes fighting over it
As of this week, four vendors are selling Gorgon Halo boxes internationally:
- GMKtec Evo-X5 Pro — $6,599 early bird ($6,799 standard) for 192GB/2TB; $6,899 for 4TB. On sale since September 28, and the first 100 units are already gone. Three M.2 slots (up to 24TB), dual 10GbE, two 80Gbps USB4 V2 ports, vapor chamber plus three fans, 120W sustained.
- Acemagic F9A Pro 495 — ~$6,499. The undercutter: same chip and 192GB, priced to steal GMKtec's launch thunder.
- Minisforum MS-S1 MAX-P495 — $7,399. The expensive one: 2TB SSD, dual 10GbE, a full-length PCIe 4.0 x4 slot, and a 120W performance mode.
- Framework Desktop — $6,799, preorders open. The repairable, tinker-friendly entry.
Still China-only for now: the FEVM FAEX2 (a 1-litre CNC “cheese grater” from ~$3,219 to ~$6,497), Beelink's GTR9 Pro 495, and Peladn's YO2 495.
The 320B-parameter claim, translated
GMKtec says the Evo-X5 Pro can run 320-billion-parameter models locally. That number depends entirely on quantization — a Q4 320B model fits in ~160GB of VRAM, sure, but nobody has published independent benchmarks yet. Remember the two halves of the story: memory capacity decides whether a model fits at all, memory bandwidth decides how fast it runs. At ~273GB/s, these boxes are fine for batch-1 inference and agents; they are not training rigs.
Should you buy one?
Not as a home server. These boxes cost ten to fifteen times the IRONHOLD shortlist and idle like workstations, not appliances — which is why none of them are picks. This segment is for local-AI workstations: big-model inference, agent workloads, and GPU-heavy creative work without the cloud bill.
If that's you: the price war is real, but wait for independent reviews. First-batch firmware on boutique boxes is historically rough, and early-bird pricing here rewards patience in reverse — prices are falling week to week, not rising. For most people, the sane local-AI buy right now is still the Mac Mini M6 at $779.99: a fraction of the price, dead silent, and enough for the models most people actually run.
Verdict
Gorgon Halo's real news isn't the chip — it's that 192GB of unified memory just became a commodity tier with four vendors fighting over it. Watch the prices, not the press releases. We'll keep tracking this segment from the sidelines; if one of these boxes ever makes sense at a home-server price, you'll read about it here first.