Lenovo ThinkCentre X Ultra vs. Acemagic F9A: Compact PCs Built for Bigger Local AI Models

Compact AI desktops are becoming less about a small NPU and more about whether a system has enough fast memory to hold a useful model locally. Lenovo’s ThinkCentre X Ultra and Acemagic’s F9A both use AMD’s Ryzen AI Max+ PRO 495, a processor platform that combines CPU, Radeon graphics, and an NPU around a shared LPDDR5X memory pool. AMD rates the chip at up to 131 total AI TOPS, including a 55-TOPS NPU, and supports up to 192GB of memory. 23

Why unified memory changes local LLM options

In these systems, the CPU, integrated Radeon GPU, and NPU draw from the same memory rather than requiring a separate discrete GPU with its own fixed VRAM allocation. That design lets compact PCs expose far more memory to GPU-oriented AI work than typical integrated-graphics machines. It is particularly relevant when model weights are the limiting factor for local inference.

Memory capacity is not a direct measure of response speed. A model’s usable performance will still vary with its architecture, quantization, context window, inference runtime, and the workstation’s sustained thermal behavior. But having enough memory is the first requirement: a model that cannot fit locally cannot run locally.

Lenovo ThinkCentre X Ultra: the enterprise-oriented option

Lenovo’s ThinkCentre X Ultra fits the Ryzen AI Max+ PRO 495 platform into a 1.6-liter desktop. The top configuration pairs the processor with Radeon 8065S integrated graphics and up to 128GB of unified LPDDR5X memory; Lenovo says up to 96GB can be configured as dedicated graphics memory. 1

The underlying processor has 16 Zen 5 CPU cores and 32 threads, while the Radeon 8065S integrates 40 compute units, according to reporting on the platform. 3 Lenovo positions the system for localized, multi-agent enterprise AI workloads, rather than as a conventional consumer mini PC. 2

Four systems, up to 512GB combined memory

A defining X Ultra feature is Lenovo’s option to cluster up to four units. At the maximum 128GB per system, that creates up to 512GB of aggregate unified memory and substantially more total compute for distributed AI workloads. 3

That should not be read as a single physically shared 512GB memory address space for every application. Whether an AI workload can take advantage of several machines depends on software support for distributed or model-parallel inference. The value is a vendor-supported scale-out path, not a guarantee that every local model will operate as if it were installed in one 512GB computer.

Software, ports, storage, and availability

The ThinkCentre X Ultra supports Windows 11 and Linux, and Lenovo includes the AMD Ryzen AI Developer Center with AI tools, models, and workflows. 8 Reported connectivity includes dual Thunderbolt 4 ports, 10GbE Ethernet, DisplayPort 2.1, and HDMI 2.1. 2 Configurations have been reported with up to two M.2 2280 PCIe Gen5 SSDs, each up to 4TB. 20

Lenovo lists availability beginning in November 2026, with a European starting price of €3,100. US reporting puts the starting price at $3,699. 13

Acemagic F9A: maximum model capacity in one box

Acemagic takes the same Ryzen AI Max+ PRO 495 platform in a different direction: maximize memory in a single mini workstation. The F9A supports up to 192GB of LPDDR5X-8533 unified memory, with up to 160GB assignable to graphics workloads. 1542

Like Lenovo’s top configuration, it uses a 16-core/32-thread Zen 5 processor and Radeon 8065S graphics. The platform’s NPU is rated at up to 55 TOPS and AMD rates total AI performance at up to 131 TOPS. 1523

Acemagic says the F9A’s memory capacity can support local inference of 120B-parameter Q4 models and 300B mixture-of-experts models. Those are vendor claims about model fit and potential use, not a universal performance guarantee; model architecture and software configuration matter greatly. 40

Expansion and physical footprint

The F9A measures 158.5 × 158.5 × 81.5 mm, or roughly two liters. 11 Its reported I/O includes two USB4 ports, dual 2.5GbE Ethernet, Wi-Fi 7, Bluetooth 5.4, HDMI 2.1, DisplayPort 2.1, and OCuLink over PCIe 4.0 x4 for external expansion. 41

Storage details are less consistent across published reports: some describe two M.2 PCIe 4.0 NVMe slots, while others describe a single M.2 2280 PCIe 4.0 x4 slot. Prospective buyers should confirm the final regional configuration directly with Acemagic. 4344

At the time of the IFA announcement, Acemagic had not published a confirmed retail price or launch date for the F9A with the PRO 495 and 192GB configuration. Prices attached to earlier Ryzen AI Max+ 395 versions should not be treated as pricing for this newer specification. 1147

Which system is the better fit?

Priority	Better match	Why
Enterprise deployment and scale-out AI	ThinkCentre X Ultra	Windows/Linux positioning, developer tooling, 10GbE connectivity, and a stated four-node clustering path. 28
Largest local model that can fit in one compact PC	Acemagic F9A	Up to 192GB unified memory, versus Lenovo’s 128GB ceiling. 115
External expansion	Acemagic F9A	OCuLink PCIe 4.0 x4 is explicitly part of the published connectivity. 41
Smallest chassis	ThinkCentre X Ultra	1.6 liters, compared with the F9A’s roughly 2-liter enclosure. 111

The choice is therefore less about the shared processor than the system design around it. Lenovo offers a more clearly enterprise-packaged route to scaling workloads across several compact machines. Acemagic makes the case for a single box with the largest available memory pool, which is often the cleaner approach when a model needs to fit entirely on one machine.

The growing mini-workstation AI market

These launches show how the local-AI PC category is moving beyond modest NPU ratings. For LLM work, model fit, memory bandwidth, networking, external expansion, and runtime compatibility can matter as much as headline TOPS.

NVIDIA’s RTX Spark systems are the emerging alternative. NVIDIA has said RTX Spark Windows PCs, built around a Grace CPU and Blackwell RTX GPU, will arrive in October 2026 from more than six OEM partners, with up to 128GB of unified memory and up to one petaflop of AI compute cited for the platform. 24 ASUS has announced the compact RTX Spark-based ProArt GR1X mini PC, while Acer has shown an RTX Spark small-form-factor desktop design but says availability will be announced later. 3439

The AMD systems and upcoming RTX Spark machines represent competing approaches to the same demand: putting enough compute and memory on a desk to run serious AI workloads locally. For buyers, the practical comparison will come down to the model sizes they need to fit, the inference software they rely on, their need for CUDA-specific tooling, and whether one high-memory system or several networked nodes is the better operational design.