Lenovo ThinkCentre X Ultra vs. Acemagic F9A for Local AI

AMD’s Ryzen AI Max+ PRO 495 gives these compact workstations an unusual advantage for local AI: the CPU, Radeon 8065S integrated graphics and AI resources draw from the same LPDDR5X memory pool. Instead of being constrained by a separate, fixed amount of discrete GPU VRAM, a local-model runtime can access a much larger shared-memory budget. 442

That is the central appeal of Lenovo’s ThinkCentre X Ultra and Acemagic’s F9A. They are not interchangeable, however. Lenovo is positioning its 1.6-liter system as a managed business workstation with a stated scale-out option, while Acemagic is pursuing the biggest announced single-machine memory configuration.

At a glance

	Lenovo ThinkCentre X Ultra	Acemagic F9A PRO 495
Processor	Up to Ryzen AI Max+ PRO 495 with Radeon 8065S graphics 42	Ryzen AI Max+ PRO 495 with Radeon 8065S graphics 4
Maximum unified memory	128GB LPDDR5X 42	192GB LPDDR5X-8533 4
Maximum graphics-memory allocation	Up to 96GB allocated from unified memory 42	Up to 160GB allocated to graphics, according to reported specifications 23
Form factor	1.6 liters 42	About 2 liters; 158.5 × 158.5 × 81.5 mm 5
Storage	Up to two 4TB M.2 2280 PCIe Gen5 SSDs, or 8TB total 33	One M.2 2280 PCIe 4.0 x4 slot supporting up to 4TB has been reported 13
Networking and expansion	Enterprise-oriented configuration with AMD PRO and Lenovo security/manageability features; detailed I/O varies by configuration 42	Dual 2.5GbE, Wi‑Fi 7, two USB4 ports, HDMI 2.1, DisplayPort 2.1 and OCuLink 4
Software position	Windows 11 and Linux compatibility, plus Lenovo’s AI Developer Center and business-management stack 45	Local-AI-focused hardware announcement; equivalent enterprise fleet-management commitments were not specified in the launch material 4
Availability and price	Expected from November 2026, starting at €3,100 42	Price, regions and retail date had not been announced at the IFA showcase 1

Why unified memory matters for local models

A conventional desktop with a discrete GPU separates system memory from GPU VRAM. For local inference, model weights and the growing KV cache for a long conversation must fit into the memory accessible to the accelerator running the model. That makes VRAM capacity a common limit.

The PRO 495 design replaces that hard division with shared memory. Acemagic explicitly says its CPU, integrated Radeon graphics and AI resources share the same pool, which it positions for local AI inference and other memory-intensive workloads. 4 Lenovo similarly specifies up to 128GB of unified memory, with up to 96GB allocated as graphics memory. 42

This does not turn shared LPDDR5X into the equivalent of high-end discrete GPU memory. Its value is primarily capacity and flexibility: it can enable larger quantized models, longer contexts or more KV-cache headroom than a small discrete-VRAM configuration. Actual generation speed remains dependent on the model format, inference engine, prompt/context size, memory bandwidth and sustained power behavior.

Same silicon, different memory ceiling

Both machines pair the PRO 495 with Radeon 8065S integrated graphics. Acemagic lists the top configuration with 16 CPU cores and 32 threads, up to 5.2GHz boost, and a 55-TOPS NPU; AMD’s quoted total platform figure is up to 131 TOPS. 47 Lenovo’s announced system also reaches a 55-TOPS NPU and uses the same Radeon 8065S graphics. 3342

Those shared headline specifications mean the biggest difference is not the chip—it is the amount of memory each vendor makes available.

Acemagic F9A: capacity-first local inference

The F9A’s maximum 192GB LPDDR5X-8533 configuration is its defining feature. Reported specifications say up to 160GB can be assigned to the integrated GPU. 23 That extra 64GB over Lenovo’s maximum system memory can matter substantially when a model, its context window and its KV cache compete for the same pool.

Acemagic has promoted the system for models up to 120B in Q4 quantization and 300B mixture-of-experts models. 2 These should be read as vendor fit claims, not as independent evidence of interactive performance. A model that loads successfully may still generate too slowly for a particular use case, especially with a long context or a demanding runtime.

The F9A also has enthusiast-friendly expansion: two USB4 ports, dual 2.5GbE, Wi‑Fi 7, display outputs and OCuLink. 4 OCuLink is particularly notable because it offers a route to external PCIe expansion, including an external GPU, although the usefulness of that setup depends on software and the attached hardware.

Lenovo ThinkCentre X Ultra: a more defined enterprise path

Lenovo’s X Ultra tops out at 128GB of unified memory and up to 96GB of allocated graphics memory. 42 That is less headroom than Acemagic’s maximum configuration, but it remains a large pool for a desktop of this size.

Its differentiator is a stated cluster option: Lenovo says an X Ultra can connect to as many as three additional units for larger models and multi-agent workloads. 44 Four fully configured systems would provide up to 512GB of aggregate unified memory.

That aggregate figure should not be mistaken for one coherent 512GB GPU-memory space. Multi-node inference requires software that can divide work or model layers across systems, and network communication adds overhead. The cluster may expand what can be run locally, but it does not guarantee the latency or simplicity of a single accelerator with the same amount of directly attached memory.

Lenovo also supplies a clearer business platform around the hardware, including AMD PRO support, ThinkShield security and DASH manageability. 36 It lists Windows 11 and Linux compatibility, and its expected November 2026 launch and €3,100 starting price give prospective buyers a more concrete procurement target. 4245

What the AI-performance numbers do—and do not—say

The 55-TOPS NPU rating and up-to-131-TOPS platform figure are useful indicators of theoretical supported AI capability, but they are not a benchmark for chatbot response speed or model quality. Neither announcement supplies independent, comparable tokens-per-second results across common LLMs.

For a buying decision, separate three questions:

1. Will the model fit? Memory capacity, quantization and context length dominate this question.
2. Will it be responsive? This requires model- and runtime-specific tokens-per-second testing, not a TOPS figure alone.
3. Can it be operated reliably at scale? Here Lenovo’s management tooling and documented multi-system route may matter more than peak single-node capacity.

For one large quantized model in a compact box, the F9A’s 192GB maximum is the more ambitious specification. For an organization that values a known launch timetable, Linux/Windows options, management features and a vendor-described cluster setup, Lenovo’s X Ultra is the more structured proposition.

How RTX Spark changes the compact local-AI market

The two AMD systems arrive as Nvidia is expanding RTX Spark into Windows laptops and compact desktops. Nvidia said RTX Spark systems from ASUS, Dell, HP, Lenovo, Microsoft Surface and MSI would be available in fall 2026, with Acer and Gigabyte systems to follow. 32 Reuters likewise reported that the platform would appear in laptops and compact desktops from major OEMs, including Dell, HP and Lenovo. 17

RTX Spark is therefore a meaningful competitive direction, especially for buyers who prioritize Nvidia’s GPU software ecosystem. But it is not yet a direct, model-for-model substitute for either machine: final OEM configurations, pricing and independent performance data vary by device and still need validation. 23

The practical choice

• Choose the Acemagic F9A if the priority is maximum announced single-box unified memory. Its 192GB ceiling and reported 160GB graphics allocation are the clearest advantages for experimenting with bigger quantized models and longer contexts. Its purchase details remained unconfirmed at announcement. 12
• Choose the Lenovo ThinkCentre X Ultra if business manageability, supported operating-system choices and a formal path to a four-system cluster are more important than maximum memory per machine. It is expected from November 2026, starting at €3,100. 4245
• Wait for benchmark comparisons if interactive LLM speed is the deciding factor. Memory capacity makes these systems unusual, but no announced TOPS number can replace tests using the exact model, quantization, context length and inference software you plan to run.

The broader takeaway is that compact AI workstations are beginning to compete on memory capacity as much as raw accelerator throughput. In this segment, the ability to keep a model local—and keep enough memory free for useful context—can be more important than a headline compute number.