AMD has put a date and a price on its Ryzen AI Halo, the compact AI workstation it teased at CES in January. Pre-orders open in June 2026 with pricing starting at $3,999 (€3,680), and the box ships with full Linux support out of the gate alongside Windows.
Inside the 15.0 cm x 15.0 cm x 4.3 cm (5.9 x 5.9 x 1.7 inches) chassis sits the Ryzen AI Max+ 395, a 16-core, 32-thread Zen 5 part paired with a 40-core Radeon 8060S iGPU built on RDNA 3.5. The headline spec is 128GB of LPDDR5x-8000 soldered as unified memory, which the integrated graphics can address directly for local model inference. Storage tops out at 2TB of PCIe 4.0 x4 NVMe, networking covers WiFi 7, Bluetooth 5.4, and 10 Gbps Ethernet, and the I/O includes HDMI 2.1b alongside three USB Type-C ports plus a separate USB-C power input.
AMD is aiming this squarely at the NVIDIA DGX Spark, and the dual-OS story is the pitch: where Spark is Linux-only on NVIDIA's ARM-based stack, the Ryzen AI Halo runs the standard x86_64 Linux distributions developers already use for local LLM work, llama.cpp builds, ROCm experimentation, or just as a general workstation when it isn't crunching tokens. The 128GB pool is large enough to hold quantized 70B-class models entirely in memory without paging to disk.
For the open-source inference stack, AMD is shipping validated environments for vLLM, Ollama, and ComfyUI alongside llama.cpp, and the ROCm 7.2 release extends compute compatibility across both Linux and Windows. On the kernel side, AMD has already submitted amd_halo_led, an upstream x86 platform driver targeting Linux 7.2 for the unit's RGB LED light bar, following the standard upstream contribution path rather than keeping device-specific code in a vendor branch. Community tooling is forming in parallel, with the amd-strix-halo-toolboxes project on GitHub already offering Docker images pre-built for the gfx1151 GPU architecture.
A second SKU is on the roadmap built around the Ryzen AI Max+ PRO 495, pushing memory capacity to 192GB with up to 160GB addressable as video RAM and modest bumps to CPU and GPU boost clocks. It stays on the same Zen 5 and RDNA 3.5 silicon, so existing software stacks should carry over without changes when that variant lands.
