AINEWS
AMD's top Ryzen AI Max chip powers a small desktop whose 192GB of unified memory is what lets it run large language models locally, not the peak AI throughput number on AMD's spec sheet.
For local AI inference, the binding constraint is rarely throughput. It's memory. A 70-billion-parameter large language model in 8-bit quantization weighs around 70GB just for the weights; 16-bit doubles that, and a useful context window adds gigabytes more before the first token prints. Framework's previewed 4.5-liter mini-ITX desktop with 192GB of unified LPDDR5X is the first off-the-shelf machine built around that fact, and it makes the case more directly than any "AI PC" marketing slide to date.
The headline spec is the memory pool, not the silicon. Framework's product page lists 192GB of LPDDR5X at 8,533 MT/s, which works out to 273 GB/s of bandwidth shared between the CPU and the integrated Radeon 8065S graphics. For comparison, a typical consumer desktop ships with 16–32GB of DDR5, and a workstation AI box usually tops out at 64–128GB. Capacity is the difference between fitting a 70B-class model and a 7B-class one; bandwidth is the difference between tokens-per-second that feel like a real assistant and ones that feel like a demo.
The chip driving it is AMD's Ryzen AI Max+ PRO 495, the stepped SKU over the existing Max+ 395: 16 Zen 5 cores and 32 threads at a 5.2 GHz boost, the same 40 compute units of RDNA 3.5 graphics, but with the integrated GPU clocked up to 3.0 GHz. AMD rates the silicon at 131 TOPS of AI compute. On a unified-memory APU, the iGPU pulls from the same 273 GB/s pool as the CPU, so the constraint on which models you can run is how much of the 192GB you can devote to weights, not the peak TOPS the marketing brochure quotes.
Framework's product page makes the same argument. The company previewed the box on X and on the same page as a host for local LLM inference, and it claims the 192GB configuration can run DeepSeek V4-Flash at Q8 quantization "with room to spare for context length." That claim is a Framework-sourced capability statement, not an independent benchmark; third-party press, including Notebookcheck and Wccftech, repeats the framing rather than running its own test. The 192GB box is also a 50% capacity bump and a 6.6% bandwidth bump over the existing 128GB Framework Desktop, in the same 4.5L chassis.
First, an open-ended PCIe x4 slot, the only listed I/O difference for the 192GB SKU, lets the box host longer expansion cards than the standard layout. VideoCardz notes that no other chassis or cooling changes are documented. Second, Framework showed a proof-of-concept 2-node cluster that pairs two of the new desktops with dual 50GbE NICs running RDMA over Ethernet (RoCE) to pool 384GB across both machines via tensor parallelism. That setup is explicitly a demonstration, not a shipping product feature; calling it anything more would overstate what the page sells.
Whether the 192GB box is a category shift or an expensive curiosity turns on the one number the page does not publish: the price. The 192GB SKU is listed as "coming soon," and Framework has warned the new configuration will carry a substantial premium over the 128GB model because LPDDR5X pricing is rising. Fudzilla pegs the current 128GB Framework Desktop at roughly $3,500 to start, and VideoCardz confirms the 192GB pricing is unannounced. A 50% memory bump on a tier where LPDDR5X is the dominant cost line does not get cheap.
The comparison Framework is betting on is not the $1,500 consumer desktop with 32GB of RAM and a discrete GPU, but the cloud endpoint charging by the hour for the same model. If the 192GB box lands near or under the cost of running a 70B-class model in the cloud for a year, the calculus changes. If it lands well above that, the 192GB SKU is a halo product for the 128GB line that ships today. Framework's product page is the variable; the next move is the price tag.