THE CRUNCH
AMD has held a press briefing on its "Agentic PC" push, a company term covering PCs whose CPU, GPU and NPU can run AI agents locally. The company says it has shipped over half a million Agentic PCs across more than 50 designs, alongside tens of millions of AI PCs across more than 250 designs. The briefing focused on the Ryzen AI Max 400 series, pitched as the "next-gen Agentic PC" for 2026, and particularly the flagship Ryzen AI Max+ PRO 495 "Gorgon Halo" with up to 192 GB of unified LPDDR5X memory, part of which can be manually assigned to the GPU.
AMD says the chip runs Z.ai's GLM 5.3 Flash, a 320-billion-parameter model that took roughly 160 GB at 4-bit quantization, at up to 20 tokens per second, and Qwen 3.8 Flash Next at up to 42 tokens per second with MTP enabled. Slides headline "300 billion parameter support", though an endnote still referencing 200B-parameter models for the 128 GB configuration appears to be a leftover. AMD also calls it the only Windows PC that can run "Opus 4.8 class" agents.
In ComfyUI comparisons, AMD showed the chip ahead of Intel's Core Ultra X9 388H in all 18 generative AI workloads, with 17 leads between 1.1x and 3.1x. The outlier Yuve result of 32.2x likely reflects memory capacity: AMD's system had 192 GB with 128 GB assigned to the GPU, while the Intel chip tops out at 96 GB of LPDDR5X, and no Panther Lake laptop on sale ships with more than 64 GB. AMD also misspelled several model names in the chart.
AMD leaned on x86 compatibility, saying Windows and Linux run natively without emulation, unlike Arm-based Snapdragon PCs that rely on Microsoft's Prism, which matters for professional apps such as Solidworks and Autodesk 3ds Max. Partnerships include Microsoft picking Ryzen AI Halo as the first Project Zenith platform, Perplexity, and Cisco. A Light Sail VR case study claims a local AI production coordinator let the studio go from one or two jobs at a time to six to nine, though those claims and all performance figures come from AMD's own testing and have not been independently verified.