AMD launches Ryzen AI Halo: mini PC for AI that competes with DGX Spark, running both Windows and Linux

This is a platform for AI developers, in the form of mini PCs using the AMD Ryzen AI Max+ "Strix Halo" platform, which is AMD's answer to NVIDIA DGX Spark as well as Apple's Mac Studio. AMD Ryzen AI Max+ is not new, but AMD Ryzen AI Halo will officially launch in Q2/2026.

You can consider these as AI workstations, serving on-premises AI needs with the ability to run models with up to 200B parameters, using unified memory, 60 FP16 TFLOPS GPU, up to 50 TOPS NPU, optimized for local AI inference, not dependent on the cloud.

Ryzen AI Halo: on-premises AI platform for developers

AMD introduces Ryzen AI Halo as an AIO solution for developing and running AI locally, aimed at those working with large models and complex AI workflows. The product is designed to reduce system configuration time, helping users go from concept to operational pipeline in just minutes after turning on the machine. The good thing about Ryzen AI Halo is that it can run both Linux and Windows, and is fully supported by drivers and necessary frameworks from AMD like ROCm, instead of "Linux only" like NVIDIA's DGX Spark or Mac Studio's macOS. The device comes pre-installed with the Ryzen AI Halo Developer Center application, which acts as a hub to access playbooks, tools and updates, and supports frameworks and tools such as LM Studio, Ollama, PyTorch, vLLM, LoRA and many other software for AI programmers.

Performance and hardware configuration

Ryzen AI Halo uses the Ryzen AI Max+ 395 platform, comes with 128GB of LPDDR5x unified memory, allowing processing and storing large model data on the same common memory space for CPU, GPU and NPU. The internal GPU reaches up to 60–16 TFLOPS, while the NPU can max out at around 50 TOPS, targeting hardware-accelerated AI workloads. AMD also mentioned that the new Ryzen AI Halo version will use Ryzen AI Max+ PRO 495 with memory support up to 192GB, expanding the ability to run large models. The company also highlights the ability to maintain continuous AI inference with low one-time hardware costs and monthly electricity costs, with simulated usage scenarios showing that the total three-year cost can be up to 6 times lower than using cloud APIs with equivalent load.

AMD released some internal benchmark results, showing that Ryzen AI Halo can run about 3.3 times to 7.3 times faster than Apple M4 Pro depending on the model and task, for example Ace Step 1.5, Ace Step 1.5 XL or Stable Diffusion XL. In other tests, the Ryzen AI Halo was compared to the NVIDIA DGX Spark system, with some models such as the GLM 4.7 Flash‑30B‑A3B or Qwen 3.5‑122B‑A10B delivering 4% to 14% higher tokens/second in scenarios published by AMD. AMD also offers the AMD AI Playbooks as learning paths and configuration guides, making it easy for new users to get started with workflows from imaging and code assistant to automation and LLM and the AMD AI Developer Program, which offers beginner to advanced support, including priority support via GitHub and official channels.