AMD Ryzen AI Halo Launch: $3,999 128GB Mini-PC Takes On Nvidia

The AMD Ryzen AI Halo launch is AMD’s clearest shot yet at the desktop AI market, and it arrives at a price that makes Nvidia’s DGX Spark look expensive. Built around a Ryzen AI Max+ 395 chip and 128GB of unified memory, this mini-PC is designed to run large language models locally — no cloud bill, no rented GPU. Here’s what shipped, what it costs, and whether it really beats the competition.

At its core, the Ryzen AI Halo is a compact developer box — roughly 150 x 150 x 43mm and just over a kilogram — aimed squarely at people who build and run AI models. AMD positions it as a turnkey workstation for local inference, not a gaming rig or a general-purpose desktop. In other words, it is a purpose-built AI appliance you keep on your desk.

However, the headline is the memory. With 128GB of unified LPDDR5X-8000 shared between the CPU, GPU and NPU, the machine can hold models that simply will not fit on a typical consumer graphics card. That single design choice is what makes the whole pitch work.

Driving everything is the Ryzen AI Max+ 395, AMD’s Strix Halo silicon. It pairs up to 16 Zen 5 cores and 32 threads, boosting to 5.1GHz, with a 40-compute-unit Radeon 8060S GPU on the RDNA 3.5 architecture, plus a dedicated XDNA 2 NPU for AI acceleration. As a result, three different engines share the same giant memory pool.

The rest of the spec sheet is deliberately practical. You get a 2TB PCIe Gen4 SSD, HDMI 2.1, four USB-C ports, 10GbE networking, Wi-Fi 7 and Bluetooth 5.4. Meanwhile, the soldered memory keeps bandwidth high — a necessary trade-off when you are feeding a 40-core GPU and an NPU at once.

This is where the AMD Ryzen AI Halo launch gets interesting. AMD set the price at $3,999 — the exact figure Nvidia’s DGX Spark debuted at back in October 2025. Since then, though, Nvidia’s box has crept up to $4,679 because of LPDDR5X and NAND supply constraints.

As a result, AMD now undercuts its closest rival by roughly $680 on identical-sounding 128GB hardware. Because the two machines chase the same buyer — a developer who wants a self-contained AI workstation — that gap is hard to ignore. For many small teams, $680 is the difference between one unit and two.

Price aside, the Ryzen AI Halo has a quieter advantage: it runs both Windows 11 and Linux. Nvidia’s DGX Spark, by contrast, is Linux-only. For developers who live in Windows tooling, that flexibility removes a real friction point.

It also changes the ownership story. Instead of renting cloud GPUs by the hour, you keep the model, the data and the machine in-house. Consequently, sensitive workloads never leave your desk — an appealing pitch for anyone nervous about sending prompts to a third-party API.

Hardware is only half of any AI launch, so AMD bundled a software stack to match. The Ryzen AI Halo ships with the AMD Ryzen AI Development Center, a package manager that installs pre-validated AI software tuned for the chip. In short, it tries to spare you the usual dependency nightmare.

On top of that, the system arrives with five preloaded AI playbooks, with another ten available online. These are essentially ready-to-run recipes for common jobs like local chat, retrieval and fine-tuning. For a first-time local-AI buyer, that hand-holding lowers the barrier considerably.

So who should care about the AMD Ryzen AI Halo launch? The obvious audience is AI developers and small teams who want to prototype with large models without a recurring cloud bill. With 128GB on tap, they can experiment with models far bigger than a 24GB consumer GPU allows.

Beyond that, privacy-sensitive shops — law firms, clinics, studios working under NDA — get a compelling option. Because the data stays local, compliance gets simpler. Gamers and mainstream users, on the other hand, are not the target here; this is a tool, not a toy.

AMD is not stopping at 128GB. A follow-up model built on the Ryzen AI Max+ PRO 495 is slated for Q3 2026, pushing capacity up to 192GB and, according to AMD, enabling models as large as 300 billion parameters locally. Pricing has not been confirmed yet.

If that ships on schedule, the gap to cloud-class hardware narrows even further. For now, though, the message of this launch is simple: serious local AI no longer demands a server rack or a five-figure budget.

If a 128GB mini-PC has you rethinking your setup, see how dedicated cards stack up in our guide to the best GPUs for AI and local LLMs. And if you plan to actually build on this hardware, our roundup of the best vibe coding tools pairs nicely with a local-first workflow.

*Sources: Tom’s Hardware, ServeTheHome, Wccftech, TweakTown, VideoCardz.*