AnswersPublished28 sources
GMKtec EVO-X5 Pro and Acemagic F9A Bring 192GB Local AI to Mini PCs
GMKtec and Acemagic unveiled compact Ryzen AI Max+ PRO 495 systems with up to 192GB LPDDR5X 8533 unified memory and up to 160GB assignable to graphics. The key upgrade over prior 128GB Ryzen AI Max systems is a 50% memory capacity increase, plus theoretical bandwidth of about 273GB/s versus 256GB/s for LPDDR5X 8000...
Published byEdited with GPT-5.6 TerraImages generated with GPT Image 2
Research answer
Key takeaways
- GMKtec and Acemagic unveiled compact Ryzen AI Max+ PRO 495 systems with up to 192GB LPDDR5X 8533 unified memory and up to 160GB assignable to graphics.
- The key upgrade over prior 128GB Ryzen AI Max systems is a 50% memory capacity increase, plus theoretical bandwidth of about 273GB/s versus 256GB/s for LPDDR5X 8000 implementations.
- GMKtec said EVO X5 Pro sales would begin in October 2026. Acemagic’s 192GB/2TB F9A configuration was reported at €8,000, a planned list price rather than a verified street price.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What did GMKtec and Acemagic unveil at IFA 2026 in Berlin with their EVO-X5 Pro and F9A mini workstations based on AMD’s 16-core/32-thread R. Article summary: GMKtec and Acemagic used IFA 2026 to introduce unusually memory-dense mini workstations: the EVO-X5 Pro and F9A, respectively, based on AMD’s Ryzen AI Max+ PRO 495 (“Gorgon Halo”). Their key advantage is not discrete-GPU. Topic tags: general, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fa
GMKtec’s EVO-X5 Pro and Acemagic’s F9A are among the first compact workstations built around AMD’s Ryzen AI Max+ PRO 495, formerly codenamed Gorgon Halo. Their defining feature is not a discrete GPU: it is a 192GB pool of high-bandwidth memory shared by the CPU and integrated GPU. That unusually large local pool is intended for developers who need to load heavily quantized large language models without sending prompts or documents to a cloud service. 1
8
43
The hardware unveiled at IFA 2026
The Ryzen AI Max+ PRO 495 pairs 16 Zen 5 CPU cores and 32 threads, boosting up to 5.2GHz, with 40-CU Radeon 8065S integrated graphics. AMD lists support for up to 192GB of 256-bit LPDDR5X-8533 memory; the processor’s NPU is rated for up to 55 TOPS, while vendor material cites up to 131 total AI TOPS for the platform. 1
3
43
The memory is shared rather than split between system RAM and a separate graphics card. In supported configurations, up to 160GB of the 192GB pool can be allocated to the integrated GPU as graphics memory. LPDDR5X-8533 on the 256-bit interface works out to roughly 273GB/s of theoretical bandwidth. 3
35
43
For the two IFA machines:
- GMKtec EVO-X5 Pro: GMKtec confirmed a Ryzen AI Max+ PRO 495 mini PC with up to 192GB memory and up to 160GB assignable to the GPU. The company said sales would start in October 2026.
8
- Acemagic F9A: Acemagic showed its compact F9A PRO 495 at IFA with the same processor family, up to 192GB unified memory, and Radeon 8065S graphics. One report describes a roughly two-liter chassis measuring 158.5 × 158.5 × 81.5mm.
1
29
Published reporting does not yet provide equally complete, final documentation for the two systems’ ports, cooling behavior, power limits, or regional configurations. Those details matter: sustained AI performance in a mini PC depends heavily on cooling and configured power, not just the chip name.
Why 192GB matters for 300B-parameter LLMs
A parameter count alone does not determine whether a model fits; its quantization format and runtime overhead do. For model weights alone, a 300-billion-parameter model requires approximately:
| Weight format | Approximate raw weight size | Practical fit in a 160GB GPU allocation? |
|---|---|---|
| FP16 | 600GB | No |
| INT8 | 300GB | No |
| INT4 | 150GB decimal | Potentially, with limited headroom |
At 4 bits per parameter, 300 billion weights occupy roughly 150GB before quantization metadata, runtime buffers, and the key-value cache needed for context. That is why AMD’s 300B-class positioning depends on aggressive quantization and workload configuration. AMD’s own platform messaging similarly qualifies the claim by model architecture, quantization, and workload. 15
31
The practical conclusion is simple: these systems may be able to load and run some roughly 300B-class INT4 models locally, but that is not equivalent to running a 300B model at FP16 or INT8, nor does it promise high token-generation throughput. The memory allocation also leaves limited capacity for the operating system and CPU-side work.
The cited material does not provide reliable primary evidence for a demonstration or model called “Qwen3.8-Flash-Next.” It should not be treated as an independently verified benchmark or product demonstration without a primary source.
Capacity is the real generational upgrade
Compared with the prior Ryzen AI Max 300 class, which topped out at 128GB of LPDDR5X-8000 unified memory, the PRO 400 platform raises the ceiling to 192GB—an increase of 50%. The maximum graphics-memory allocation rises from 96GB to 160GB, according to reporting on AMD’s platform announcement. 11
Bandwidth rises more modestly. A 256-bit LPDDR5X-8000 setup provides about 256GB/s of theoretical bandwidth, while LPDDR5X-8533 reaches roughly 273GB/s—about a 6.6% increase. The CPU remains a 16-core Zen 5 design and the flagship retains Radeon 8065S graphics, so this is principally a memory-capacity refresh rather than a wholesale compute-architecture change. 9
33
36
That distinction matters for buyers. The extra memory changes which models can stay resident locally; it does not turn an integrated GPU into a high-bandwidth HBM-equipped datacenter accelerator.
The wider 192GB Gorgon Halo ecosystem
The PRO 400 family includes the Ryzen AI Max PRO 485, 490, and Max+ PRO 495. The 495 is the top configuration, with 16 cores and Radeon 8065S graphics; the lower 485 and 490 models use Radeon 8050S graphics. 16
32
Other systems announced around this platform include:
- Minisforum MS-S1 MAX P495, a mini workstation advertised with up to 192GB LPDDR5X unified memory and up to 160GB allocatable as VRAM.
3
4
- Framework Desktop, a compact desktop configuration with the PRO 495, 192GB memory, 273GB/s stated bandwidth, and up to 160GB available to the iGPU.
14
33
- Lenovo ThinkCentre X Ultra, which was reported with up to the PRO 495 but a lower 128GB memory ceiling and up to 96GB dedicated graphics memory. It was expected to start at $3,669 and ship in November 2026.
34
Who should consider these systems?
A 192GB Gorgon Halo mini workstation is best understood as a local-inference and development machine for workloads where model capacity and data locality matter more than maximum throughput. It may appeal to teams experimenting with quantized LLMs, private document workflows, or offline AI development that would otherwise require a cloud instance or a much larger multi-GPU workstation.
There are clear limits. A shared-memory Radeon iGPU with roughly 273GB/s of bandwidth is very different from a high-end accelerator with HBM memory and substantially greater bandwidth. The NPU’s 55-TOPS rating is also not a direct measure of large-LLM generation performance; software stack, GPU execution path, model format, context length, and sustained power all affect real results. 1
35
The cost is another constraint. Acemagic was reported to be planning a 192GB/2TB F9A for €8,000 through its own store and Amazon in September 2026. Treat that as a reported planned list price, not a confirmed retail transaction price. 6
Still, the EVO-X5 Pro and F9A show where compact AI workstations are heading: toward systems built around enough unified memory to keep much larger quantized models on one desk-bound machine. For buyers, the most useful question is not simply whether a machine claims “300B” support, but which specific quantized model and context length it can run at an acceptable speed.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "GMKtec EVO-X5 Pro and Acemagic F9A Bring 192GB Local AI to Mini PCs"?
GMKtec and Acemagic unveiled compact Ryzen AI Max+ PRO 495 systems with up to 192GB LPDDR5X 8533 unified memory and up to 160GB assignable to graphics.
What are the key points to validate first?
GMKtec and Acemagic unveiled compact Ryzen AI Max+ PRO 495 systems with up to 192GB LPDDR5X 8533 unified memory and up to 160GB assignable to graphics. The key upgrade over prior 128GB Ryzen AI Max systems is a 50% memory capacity increase, plus theoretical bandwidth of about 273GB/s versus 256GB/s for LPDDR5X 8000 implementations.
What should I do next in practice?
GMKtec said EVO X5 Pro sales would begin in October 2026. Acemagic’s 192GB/2TB F9A configuration was reported at €8,000, a planned list price rather than a verified street price.