Photo: Jacob Bobo
The AMD Ryzen AI Halo is an AI-focused desktop that takes AMD's latest AI hardware and puts it in a box that can fit on any desktop. Why would you want a device like this? Because it shifts the ability to run massive large language models, or intensive stable diffusion pipelines directly onto your desk. More importantly, this AI desktop is a dedicated piece of hardware aimed at turning those ongoing cloud AI expenses into a one-time hardware investment, thanks to its AMD Strix Halo platform and 128GB of shared memory.
The Ryzen AI Halo is a Micro Center retail exclusive.
Unboxing and setup
Inside the main box, you'll find the ultra-compact system itself, accompanied by a power adapter and a straightforward quick start guide to map out your initial setup. The chassis itself is around 1.2 kilograms (2.65 pounds) and maintains a remarkably small desktop footprint of just 150 by 150 by 45.4 millimeters.
Turning the unit around reveals a rear I/O panel built to accommodate data throughput. Network connectivity is anchored by a 10Gbps RJ-45 Ethernet port for rapid local data transfers and fast network connections. Alongside the network jack and power connection sit three high-bandwidth USB-C® ports (one USB-C® 3.2 and two USB-C® 4.0) and a single HDMI 2.1 video output to handle an external display.
On the front of the machine, a diagnostic LED strip serves as your primary window into the system's status, which should usually be a solid white light.
The Ryzen AI Halo will be available in both Linux and Windows versions. Most AI developers will choose Linux, however, our test system was the Windows version. Once you land on the operating system desktop, the AMD Ryzen AI Developer Center application appears as the central command hub for the machine. This handy collection of software and playbooks should save you time versus manually configuring a setup from scratch.
Maxing out the unified memory
Under the hood, the system is driven by the AMD Ryzen AI Max+ 395 processor, a piece of silicon containing 16 high-performance Zen 5 CPU cores, 40 RDNA 3.5 graphics compute units, and an advanced XDNA 2 neural processing unit capable of delivering up to 50 TOPS of dedicated AI compute.
But the true hardware crown jewel for local AI model execution is the system's 128GB of unified LPDDR5x memory, pushing up to 256GB/s bandwidth. Because this memory pool is shared across the system processing cores, managing how that capacity is split between traditional system tasks and video memory dictates how large of a local model you can comfortably run (don't forget that context length and KV cache are as important as model size).
Right out of the box, you can configure that split with 96GB of that total unified memory explicitly as dedicated VRAM, leaving a comfortable 32GB behind to handle standard operating system and application runtime overhead. That massive 96GB allocation means you can instantly load massive, high-parameter models like GPT-OSS-120b or run intensive multi-model stable diffusion tasks entirely within memory. You can also dial the allocation up or down manually.
Chatting with large language models in LM Studio
For my first real performance test, I wanted to see how the system handled large language models using LM Studio, which provides a clean, consumer-friendly chat window right on the desktop. I opened up the Developer Center app and launched LM Studio, then headed straight to the built-in search tab to find some models. The platform gave me immediate access to download massive open-source files like the 59-gigabyte GPT-OSS-120b model, as well as efficient coding assistants like Qwen3-Coder-30B.
After my chosen models finished downloading (some came pre-loaded on my test system), AMD suggested some settings tweaks, including bumping up the Context Length setting, which gives the model a longer memory for in-depth conversations and pushing the GPU Offload all the way to its maximum.
I ran my own in-house benchmarks, which involve giving the models I chose a word problem, a science problem, and a math/logic problem, then recording the tokens per second (TPS) and the time to first token (TTFT) in seconds for each.
Even the massive GPT-OSS-120b ran at around 45 tokens per second, which is about the same speed as many cloud-based frontier AI tools.
| Model | Run | Tokens per second (TPS) | Total tokens | Time to first token (TTFT) |
|---|---|---|---|---|
| GPT-OSS-120b | 1 | 45.09 | 383 | 0.96 |
| 2 | 45.18 | 448 | 0.51 | |
| 3 | 45.12 | 633 | 0.54 | |
| Gemma-4-31b | 1 | 11.63 | 636 | 0.71 |
| 2 | 11.40 | 902 | 0.45 | |
| 3 | 11.74 | 328 | 0.60 |
Generating local art with ComfyUI
Next, I wanted to see how the system handled offline image generation using the pre-installed ComfyUI playbook. The great thing about running this locally is that you don't have to worry about paying per token or per generation, or running out of image credits -- everything is running locally.
Photo: Dan Ackerman
I fired up the ComfyUI Launcher directly from the developer dashboard and the system already had the Z Image Turbo diffusion model ready to go. I typed a descriptive visual prompt into the default template and hit generate. Because the system keeps all of these distinct model elements inside that 96GB graphics memory buffer, the images generated within a few seconds. And unlike cloud-based image gen tools, I didn't have to worry about running out of generative credits.
Chatting via Open WebUI
While LM Studio is fantastic for a quick desktop chat, I also wanted to try an environment that feels more like using ChatGPT or Claude, but running completely offline. For that, I turned to the Open WebUI playbook. Open WebUI gives you a browser-based interface to interact with your local large language models, and it can even be accessed securely by other systems on your home Wi-Fi network.
Following the playbook instructions, I utilized the pre-installed Lemonade Server running on the machine to host the LLM backend locally, and then hooked Open WebUI directly up to it. There, I was able to access a very familiar-feeling interface and organize different conversation threads seamlessly. And even better, because the service runs as a network host, I simply typed the Halo's local IP address followed by :8080 into my browser on a separate PC on the same network, effectively turning the compact device into a headless AI server.
Just the beginning
This first look at unboxing and spinning up the AMD Ryzen AI Halo only scratches the surface of what this compact machine can handle. Successfully getting my initial chat answers in LM Studio and watching images render almost instantly in ComfyUI proved that high-end local AI is a reality, but this introductory hands-on is just step one, so check back for more deep-dive testing and benchmarking still ahead.
More from MC News
- Run AI Locally: The Best LLMs for 8GB, 16GB, 32GB Memory and Beyond
- Quantization Explained: Why the Same LLM Gives Better Results on High-End Hardware
- Why Context Length, Not AI Model Size, Is Your Biggest VRAM Problem
- How to Use LM Studio to turn your PC into an AI Server
- Why VRAM and Memory Bandwidth are Key for Powering Local AI
- Keyboard 101: Intro to Computer Keyboards
- Fix It Yourself: Talking to iFixit on Why Repairable Tech Matters
- The Local LLM Your Laptop's Been Waiting For
- Gemma vs Qwen: Which Local AI is Best?
Micro Center Editor-in-Chief Dan Ackerman is a veteran technology journalist with over 20 years of experience testing and reviewing the latest consumer tech. He previously served as Editor-in-Chief of Gizmodo and Editorial Director at CNET and is the author of the New York Times-reviewed book The Tetris Effect, a critically acclaimed Cold War history of the world's most influential video game. Reach him at dackerman at microcenter.com.