AMD unveiled its Ryzen AI Max PRO 400 series this week, a three-chip refresh of the Strix Halo platform aimed at workstation-class laptops and small desktops. The headline number is memory: up to 192GB of unified LPDDR5X on a 256-bit bus, with as much as 160GB carve-out-able as VRAM for the integrated GPU. AMD says that's enough to run a 300 billion parameter model on a single x86 client chip, a first for the category.
That memory ceiling is the whole story here, and AMD knows it.
Same chip, bigger wallet for RAM
Look past the memory and the 400 series is the 300 series with a new sticker. The flagship 16-core Ryzen AI Max+ PRO 495 boosts to 5.2 GHz and pairs with a 40-compute-unit RDNA 3.5 GPU, the same Radeon configuration AMD has shipped for over a year. The one genuine hardware change: that GPU now clocks to 3.0 GHz, a 100 MHz bump over the outgoing Max+ 395. Tom's Hardware put it bluntly, noting you could scratch off the 4 and write in a 3.
The two cheaper parts are even more familiar. The 12-core PRO 490 and 8-core PRO 485 both drop to a 32-CU Radeon 8050S and top out at 50 TOPS on the NPU rather than the 495's 55, per the spec sheet AMD confirmed. They're identical to their 385 and 390 predecessors except for the memory controller. Which is fine. A mid-cycle refresh doesn't owe anyone a new architecture. But calling it a 400 series invites comparison to a real generational leap, and this isn't one.
About that 300 billion parameter claim
Here's the part worth poking at. AMD's official announcement calls the Max+ PRO 495 "the world's first x86 client processor capable of running 300 billion parameter models locally," with a footnote doing some heavy lifting. The current 128GB Ryzen AI Halo box, running last year's silicon, manages roughly 200B parameters at 4-bit precision, according to The Register. So the 300B figure almost certainly assumes the same aggressive quantization. Run something at full precision and the math gets a lot less impressive a lot faster.
And the timing is rough. AMD is pushing memory capacity higher right as DRAM prices climb on the back of an AI-driven shortage. A 192GB machine sounds great until you price the modules. The company hasn't said what these systems will cost, though the existing 128GB Halo workstation starts at $3,999, so the ceiling models won't be cheap.
Still, the pitch makes sense for the audience AMD wants. Unified memory means the iGPU sees a pool no discrete laptop card can touch. If you're a developer who needs to fit a large model in memory and doesn't care whether tokens stream slowly, this is a genuinely useful box. Jack Huynh, SVP and GM of AMD's Computing and Graphics Group, framed it as AI moving off the cloud and onto the desk, which is the kind of thing you say when you're selling a $4,000 desk computer. The framing isn't wrong. It's just selling something.
Where are the consumer chips?
One thing AMD didn't announce: anything for regular buyers. Every SKU here carries the PRO tag, meaning commercial systems with enterprise security and management features. Reports of consumer Gorgon Halo parts arriving "later this year" are floating around, but AMD wouldn't confirm a non-PRO lineup when asked. So if you were hoping to drop one of these into a gaming handheld, wait and see.
The competitive target is obvious enough. AMD spent part of the briefing positioning this against Nvidia's DGX Spark, and it's also trying to get ahead of Nvidia's long-rumored laptop CPUs before they ship. Beating a competitor to market with last year's chip plus more RAM is a strategy. Whether it holds once Panther Lake and Nvidia's parts land is a different question.
Systems from HP, Lenovo and ASUS are slated for the third quarter of 2026. The Ryzen AI Halo developer box, still running the older 395, opens for pre-orders in June at $3,999, with the 192GB Gorgon Halo version following. No word yet on what that one costs, and given where memory prices are headed, that silence is probably deliberate.
