AMD ships Ryzen AI Halo dev kit for local LLMs

AMD has begun shipping the Ryzen AI Halo development kit, a mini workstation for AI developers priced at approximately ¥640,000 (around $3,936). The system includes 128GB of unified memory and is designed to run large language models with up to 120 billion parameters locally, reducing dependence on cloud services. It is available in Windows 11 Pro and Linux versions.

The kit is built around the Ryzen AI Max+ 395, a Strix Halo processor using Zen 5 architecture with 16 cores and 32 threads, Radeon 8060S graphics, and an XDNA 2 NPU. LTT Labs found that text generation is limited by 256GB/s memory bandwidth, with Apple’s Mac Studio using M2 Ultra or M3 Ultra reaching up to 800GB/s and delivering faster Gemma 4 results. AMD’s advantage is cost: comparable AMD mini PCs were cited at about $25.77 per gigabyte of memory versus $41.66 for an Apple M3 Ultra platform.

AMD is leaning on its Ryzen AI Developer Center, prevalidated configurations, and AI Playbooks to simplify local LLM deployment, code assistance, and fine-tuning. NVIDIA’s CUDA remains the more mature AI development platform, while AMD’s ROCm is still in preview and lacks full Windows support. The XDNA 2 NPU showed efficient lightweight inference, peaking at just 35W while achieving 20 tokens per second.

Competition is also building in Asia, where Samsung- and SK Hynix-backed Rebellions plans a KOSPI listing in the first or second quarter of next year. AMD is planning a Gorgon Halo chip for the third quarter of 2026 with 192GB of unified memory and support for models with up to 300 billion parameters.