LisaSchwartz

🚨BREAKING: AMD just made it possible to run a 235 billion parameter model on a desktop the size of a lunchbox. three days ago at AMD's developer event, lisa su walked on stage holding a mini PC in one hand and ran Qwen3 235B live. no cloud, data center or rented GPU. WHAT JUST HAPPENED: the chip is the AMD Ryzen AI Max+ 395, codename "Strix Halo" it is the first x86 chip where the CPU and GPU share the same memory pool up to 128GB of unified LPDDR5X memory on linux, up to 110GB of that is allocatable to the GPU that one design choice is what changes everything THE BENCHMARK: on DeepSeek R1 inference, this chip beat an nvidia RTX 5080 by more than 3x the 5080 has 16GB of VRAM the 5080 cannot fit the model the AMD chip can this is a memory bottleneck story, not a raw compute story THE TWO WAYS TO BUY IT: → AMD's first-party Ryzen AI Halo desktop $3,999 official dev box, full AMD support, day-one ROCm → GMKtec EVO-X2 mini PC $1,499 to $2,400 depending on config same exact chip the lunchbox lisa held on stage same chip different price tags and software ecosystems THE IMPORTANT CAVEAT: Qwen3 235B is a Mixture-of-Experts model 235 billion total parameters only 22 billion active per inference pass that is why it runs at ~11 tokens per second on this hardware you are moving active weights, not the full model THE NPU CAVEAT MOST SKIP: the chip ships with a 50 TOPS XDNA 2 NPU it does NOT accelerate LLM inference ollama, llama.cpp, LM Studio all route LLMs to the GPU if you are buying this for local LLM use, the NPU spec is marketing the GPU is doing the work THE BIGGER PICTURE: nvidia spent a decade convincing the industry that frontier AI required their hardware apple proved unified memory works on M-series silicon AMD just brought that architecture to x86 at consumer prices this is the first time you can run a frontier-class model on hardware you own outright local AI on this chip is owned intelligence the question is no longer "can I run this locally" it is "do I want to keep paying $440 a month when I do not have to" P.S. the community already flagged that the $1,499 box was not the one lisa held on stage. the one she held costs $3,999. same chip, different vendor.