Large local AI models often run into a memory limit before they run into a headline-compute limit. Ahead of IFA 2026, ACEMAGIC, Acer, and BOSGAME are using AMD Ryzen AI Max+ platforms to address that constraint with unusually large unified LPDDR5X pools: 128GB in Acer’s laptop and up to 192GB in the two planned mini workstations. The CPU, integrated Radeon GPU, and NPU share that memory, allowing a substantial portion to be assigned to graphics workloads rather than relying on a separate, limited VRAM pool.
2
7
26
35
Why unified memory is the story
For local inference, model weights, the context window, and runtime overhead all need memory. A large shared pool does not turn an integrated GPU into a data-center accelerator, but it can let a compact system load models that would not fit in the VRAM of many conventional desktop graphics cards.
The clearest example is ACEMAGIC’s F9A Pro: its 192GB LPDDR5X-8533 configuration reportedly permits up to 160GB to be assigned to graphics. That is the central appeal for local-model users—not simply its advertised 131 TOPS of platform AI performance.
2
4
ACEMAGIC F9A Pro: the 192GB compact workstation
ACEMAGIC’s F9A Pro is built around AMD’s Ryzen AI Max+ PRO 495, a 16-core, 32-thread chip with boost speeds up to 5.2GHz and integrated Radeon 8065S graphics with 40 RDNA 3.5 compute units at 3GHz. The platform is rated for up to 55 NPU TOPS and 131 total AI TOPS, with support for up to 192GB of LPDDR5X-8533 unified memory.
2
4
The approximately 2-liter system measures 158.5 × 158.5 × 81.5mm. Reported connectivity includes two USB4 ports, dual 2.5GbE Ethernet, Wi‑Fi 7, Bluetooth 5.4, HDMI 2.1, DisplayPort 2.1, and PCIe 4.0 x4 OCuLink for expansion. Reports differ on internal storage-slot details, so buyers should verify the final retail configuration.
4
6
12
ACEMAGIC says the system has demonstrated local operation of the 284-billion-parameter DeepSeek V4 Flash model. That should be read as a vendor demonstration of model loading and operation, not a published independent measure of throughput, latency, context capacity, thermals, or software compatibility.
4
A confirmed F9A Pro retail price and shipment date were not established in the launch material reviewed.
10
Acer Aspire G 3D 16: 128GB local AI in a creator laptop
Acer takes the same general architecture in a different direction. The 16-inch Aspire G 3D 16 uses the Ryzen AI Max+ 395 and Radeon 8060S integrated graphics, with up to 128GB of LPDDR5X-8000 memory and up to 2TB of M.2 SSD storage. Acer rates the system at up to 126 overall TOPS.
26
Its defining feature is not a mini-PC enclosure but a SpatialLabs glasses-free 3D IPS display. The panel runs at 3840 × 2400 in 2D mode and 1920 × 2400 in 3D mode, with a 120Hz refresh rate and 100% DCI-P3 coverage. Acer also lists Wi‑Fi 7, Bluetooth 5.4, and two USB4 Type-C ports.
19
23
26
Acer and third-party coverage position the system for large local AI workloads, but the available launch information does not provide a specific model-size ceiling or independently tested tokens-per-second result. Its 128GB memory option is therefore the more meaningful specification for prospective local-LLM users.
19
26
Acer had not announced availability or pricing for the Aspire G 3D 16 in the material reviewed.
24
BOSGAME M5 MAX: two paths to 192GB
BOSGAME’s forthcoming M5 MAX is another Ryzen AI Max+ PRO 495 mini workstation. The company says it will offer 128GB and 192GB unified-memory configurations, pairing the processor’s 16 cores and 32 threads with Radeon 8065S graphics running up to 3GHz.
35
The first announced configuration combines 128GB of LPDDR5X unified memory with a 2TB PCIe 4.0 SSD. BOSGAME says the 128GB version is scheduled for mid-October 2026, while the 192GB version is planned for early 2027.
34
35
BOSGAME has also disclosed dual 10GbE networking for the M5 MAX, a notable differentiator for a machine aimed at AI developers, creators, and software professionals. It has not published a complete final I/O specification or a price in the cited announcement, and it has not supplied independently verified model-size or inference-throughput results.
35
What these launches mean for local AI
These systems point to a growing category between a conventional PC and a rack-scale AI server: desk-side machines designed around memory capacity as much as CPU, GPU, or NPU compute.
- Memory capacity is becoming a workstation specification. A 128GB or 192GB shared pool can give an integrated GPU far more room for model weights than a typical iGPU configuration. The F9A Pro’s reported 160GB graphics-memory allocation shows why these systems are being marketed for local inference.
2
7
- The form factors are diverging. Acer packages 128GB capability in a display-centric creator and gaming laptop, while ACEMAGIC and BOSGAME emphasize compact, stationary workstations with expansion and higher-speed networking.
6
23
35
- TOPS alone is not a buying guide. Real local-AI performance depends on the model format and quantization, context length, runtime support, memory allocation, sustained thermals, and actual tokens per second. Vendor demonstrations show potential, but independent testing remains essential.
The practical competition in this segment will be decided by whether these machines can deliver reliable, well-supported local inference at a sensible price—not by maximum model-parameter claims alone.