AMD Zen 5 AI Workstations Pack 192GB RAM to Run 284-Billion-Parameter Models Locally

AMD Expands Ryzen AI Workstation Ecosystem with 192GB Memory Configurations

AMD has introduced its Ryzen AI Max+ PRO 495 processor across multiple compact AI workstation platforms, with manufacturers iBuyPower, Minisforum, and AMD itself releasing systems designed for local AI workloads and large language model development1

1

2

2

3

3

. The new processor combines 16 Zen 5 cores with Radeon 8065S integrated graphics and an XDNA 2 neural processing unit rated at 55 TOPS, but the defining feature across these releases is the jump to 192GB of LPDDR5X memory in higher-end configurations2

2

3

3

.

iBuyPower enters the compact AI workstation market with its Pro AMD AI Halo Mini PC at $3,999, featuring 128GB of LPDDR5X memory running at 8,000 MHz and a 1TB SSD in the base configuration1

1

. Minisforum counters with the MS-S1 MAX-P495 at $7,399, doubling down on memory with 192GB unified memory and a 2TB SSD, currently available for pre-order with shipments expected in mid-October 20262

2

. AMD's official Ryzen AI Halo developer platform matches Minisforum's 192GB configuration and ships with Linux for $7,299.99, with a Windows 11 version incorporating Microsoft's Project Zenith planned for later release3

3

.

Memory Capacity Drives Workstation Differentiation

The Ryzen AI Max+ PRO 495 processor represents a modest performance uplift over its predecessor, the Ryzen AI Max+ 395. CPU boost clocks increase by 100 MHz to 5.2 GHz, while the Radeon 8065S graphics with 40 RDNA 3.5 compute units gain 100 MHz to reach 3.0 GHz2

2

3

3

. The NPU improves from 50 TOPS to 55 TOPS, pushing the platform total from 126 TOPS to 131 TOPS2

2

.

The substantial change lies in memory support. The 400 series controller now handles up to 192GB of LPDDR5X across a 256-bit interface, with up to 160GB assignable to the GPU, compared to 128GB and 96GB on the previous generation2

2

3

3

. Memory speed also rises from 8,000 to 8,533 MT/s, delivering approximately 7% more bandwidth and bringing total memory bandwidth to 273GB/s2

2

3

3

.

This unified memory architecture allows the CPU and integrated Radeon GPU to access the same physical memory pool, eliminating data transfers between separate memory spaces and enabling large AI models to load entirely within available capacity3

3

. According to AMD, the 192GB configuration supports language models containing 300 billion parameters or more using 4-bit quantization3

3

.

Real-World Performance with 284-Billion-Parameter Models

Minisforum's testing demonstrates the practical impact of expanded memory capacity. A single MS-S1 MAX-P495 unit ran DeepSeek V4 Flash, a 284-billion-parameter model, at Q4 quantization at approximately 12 tokens per second2

2

. Two linked units processed Qwen3.5-397B at roughly 16 tokens per second, while a four-system cluster loaded the 380GB Q4_0 build of DeepSeek-R1 671B2

2

.

These figures require context. Both DeepSeek V4 Flash and Qwen3.5-397B use mixture-of-experts (MoE) architecture, which loads only specific "experts" into memory when needed rather than activating all parameters simultaneously. DeepSeek V4 Flash activates approximately 13 billion of its 284 billion parameters per token, while Qwen3.5-397B activates about 17 billion2

2

. Memory capacity determines whether a model fits at all, while active parameters and bandwidth govern inference speed, explaining how a system with roughly 273 GB/s bandwidth achieves double-digit token speeds on models this large2

2

.

Compact Form Factors and Connectivity Options

All three systems maintain compact footprints despite their capabilities. iBuyPower's Pro AMD AI Halo Mini PC measures 70 × 198 × 200 mm and runs Windows 11 Pro, with three M.2 storage positions including two 2280 slots and one supporting 2242 or 2280 drives with PCIe 4.0 connectivity1

1

. The system includes Wi-Fi 7, Bluetooth 5.4, two 10Gb Ethernet ports, and USB4 connections capable of 40 Gbps data transfer1

1

. Front panel access provides two USB 3.2 Gen 2 Type A ports, a USB4 Type C connection, an audio jack, and an SD card reader, while rear connections add DisplayPort 1.4, HDMI 2.1, another USB4 port, and additional USB ports1

1

.

AMD's official Ryzen AI Halo developer platform maintains a 150 × 150 mm footprint and weighs less than 1.2 kg3

3

. Connectivity mirrors the feature set with 10Gb Ethernet, Wi-Fi 7, Bluetooth 5.4, and USB-C, alongside 2TB of solid-state storage3

3

. The platform ships with a prepared Linux software environment including AMD ROCm support, enabling compatible AI frameworks to leverage the integrated GPU3

3

.

Market Positioning and Developer Focus

The pricing spread reflects different target audiences. iBuyPower's $3,999 entry point with 128GB memory targets professionals seeking compact AI workstation capabilities for creative, engineering, and data-intensive tasks without requiring the largest models1

1

. The system launches in Q4 2026 with configurable storage capacity options1

1

.

Minisforum's $7,399 MS-S1 MAX-P495 currently lists at 20% off a $9,249 list price, positioning the 192GB configuration as a premium option for developers and small teams requiring 200-to-300-billion-parameter mixture-of-experts models kept on-premises for privacy or latency considerations2

2

. The system retains the 320W internal power supply, six-heat-pipe dual-blower cooler, dual 10GbE, two 80Gbps USB4 v2 ports, and 2U rack support from the original MS-S1 MAX2

2

.

While iBuyPower's system includes two 10Gb Ethernet ports, the company has not announced compatible software or configuration details for distributing model workloads across multiple units as an AI cluster1

1

. That capability requires specific software stacks beyond hardware connectivity alone1

1

.

For developers working with generative AI applications and large language models, these compact AI workstations represent a shift toward accessible local inference hardware. The ability to run models that previously required multiple discrete graphics cards or cloud resources within a single unified-memory system addresses privacy, latency, and development workflow concerns. Watch for how software tooling evolves to take full advantage of the 16 Zen 5 cores and 192GB unified memory configurations, particularly as more developers experiment with running large AI models entirely on local AI workloads.