AMD's Ryzen AI Halo brings 128 GB unified memory to compact AI workstations

At a glance:

• AMD is launching the Ryzen AI Halo developer platform with the Ryzen AI Max+ 395, a 16-core mini PC that ships with 128 GB of unified memory and a 650 TOPS NPU — priced at $3,999 with pre-orders starting in June 2026.
• A follow-up wave of Ryzen AI Max PRO 400-series chips (Max+ PRO 495, Max PRO 490, Max PRO 485) will support up to 192 GB of RAM and are among the first x86 client chips capable of running 300B-parameter LLMs.
• The platform targets local AI development and prosumers, aiming to reduce cloud dependency for testing, fine-tuning, and running large language models — positioning itself against Nvidia's DGX Spark and the Apple Mac Mini M4 Pro.

What AMD is launching

On May 20, 2026, AMD unveiled the Ryzen AI Halo developer platform, a new line of compact mini PCs purpose-built for running large language models locally. The first and only SKU available at launch is the Ryzen AI Max+ 395, a mobile processor that AMD is calling an "absolute beast" for AI workloads. The chip packs 16 physical cores and 32 threads with boost clocks reaching 5.1 GHz, 80 MB of cache, AMD Radeon 8060S integrated graphics with 40 compute units, and a 650 TOPS NPU. It ships paired with 128 GB of unified memory in a single compact chassis.

AMD chose the Max+ 395 as the flagship launch chip because its combination of CPU, GPU, and NPU resources makes it uniquely suited for local LLM inference and development. The unified memory architecture means the entire 128 GB is available to both the processor and the AI accelerator without the fragmentation you get when splitting workloads across a discrete GPU and system RAM. That gives developers the headroom to load and test models that would choke a typical workstation or even a high-end gaming GPU.

The MSRP for the first Ryzen AI Halo mini PC is $3,999. Pre-orders begin in June 2026. AMD acknowledges the price is steep, attributing it to the cost of components and the niche nature of the product, but argues the device replaces the need for cloud compute cycles during development.

How the specs stack up across the 300 and 400 series

AMD has already teased the next wave of chips that will join the Halo lineup. The Ryzen AI Max PRO 400 series uses Zen 5 cores with RDNA 3.5 graphics and XDNA 2 NPUs, and some variants will support up to 192 GB of RAM — nearly 1.5× the launch SKU's memory. Here is the full comparison of the announced models:

• Max+ 395 — 16 cores / 32 threads, up to 5.1 GHz, 80 MB cache, Radeon 8060S (40 CUs), 650 TOPS NPU, 128 GB RAM, TDP 45–120 W
• Max+ PRO 495 — 16 cores / 32 threads, up to 5.1 GHz, 80 MB cache, Radeon 8065S (40 CUs), 55 TOPS NPU, 192 GB RAM, TDP 45–120 W
• Max PRO 490 — 12 cores / 24 threads, up to 5.0 GHz, 76 MB cache, Radeon 8050S (32 CUs), 50 TOPS NPU, 192 GB RAM, TDP 45–120 W
• Max PRO 485 — 8 cores / 16 threads, up to 5.0 GHz, 40 MB cache, Radeon 8050S (32 CUs), 50 TOPS NPU, 192 GB RAM, TDP 45–120 W

The jump to 192 GB of unified memory is the headline differentiator for the 400-series Halo boxes. For context, a typical discrete GPU setup for running LLMs at home comes with around 16 GB of VRAM. That means the top-end Halo SKU offers more than ten times the memory capacity, enabling the machine to load and process models in the 300B-parameter range — something AMD says is among the first x86 client chips to achieve.

Software and developer workflow

AMD is pairing the hardware with a ready-to-use software stack for both Windows and Linux. The Ryzen AI Developer Center includes preloaded apps and models, guided playbooks for common workflows, and full out-of-the-box support for AMD ROCm. ROCm brings optimized performance and state-of-the-art model support by leveraging a wider platform of validated tools, frameworks, and drivers.

The platform also includes up to 2 TB of PCIe Gen 4.0 storage, 10 Gbps networking, and Wi-Fi 7 connectivity, rounding out the specs for a machine that can sit on a desk and serve as a daily driver for AI experimentation without requiring a server-room footprint.

AMD's pitch is straightforward: reduce reliance on cloud resources for testing, fine-tuning, and running local models. With 128 GB of RAM on the launch SKU, developers can keep larger models resident in memory and iterate faster than they could by spinning up cloud instances for every experiment.

How it compares to competitors

AMD is positioning the Ryzen AI Halo directly against two well-known rivals. On one side is Nvidia's DGX Spark, which AMD claims is limited to Linux, lacks an NPU, and delivers weaker performance based on the company's internal benchmarks. On the other side is the Apple Mac Mini M4 Pro, which benefits from a tightly optimized hardware-software stack but tops out at running models larger than 100B parameters — a ceiling the Ryzen AI Halo clears comfortably.

The TDP range of 45–120 W for the Halo platform also makes it appealing from an operational-cost perspective. At a per-kWh price of $0.15, AMD estimates the monthly electricity bill for running the system around $16 — a fraction of what a persistent cloud GPU instance or an always-on AI workstation would cost.

That said, AMD freely admits it has lagged behind Nvidia's AI push and platform support. The company is clearly hoping the Halo lineup will help it cause a splash and win back developer mindshare, particularly among prosumers and small teams that cannot justify the expense of a multi-GPU server but still need to run large models locally.

What to watch next

Only one SKU is available at launch, and AMD says more models will roll out later in 2026. The Ryzen AI Max PRO 400-series chips are expected to bring lower-price points to the Halo family, potentially making the platform accessible to a broader audience. The first 300-series Halo has not even shipped yet, and the company has already disclosed the specifications for the 400-series, suggesting a rapid refresh cadence.

For the AI development community, the key question is whether AMD's ROCm ecosystem and software stack can match the maturity of Nvidia's CUDA tooling. If the Halo's unified-memory approach and large RAM capacities prove compelling in real-world benchmarks, it could shift the calculus for developers deciding between cloud spend and local hardware.

Tags: AMD, Ryzen AI Halo, LLM, local AI, mini PC