Acer Veriton RI110 puts 96GB of memory and OCuLink into a mini PC built for local AI

Acer’s Veriton RI110 looks like a mini PC, but its specification reads more like an argument about where AI workloads should run. The compact workstation combines an Intel Core Ultra X7 358H processor, Intel Arc B390 graphics, up to 96GB of LPDDR5X memory and as much as 4TB of PCIe 4.0 storage. Acer says it can work with AI models containing up to 120 billion parameters.

The key word is “hybrid”. The RI110 is not presented as a machine that eliminates cloud AI. It is designed to handle some models and agent tasks locally while escalating other work to remote services. That split can reduce latency and keep sensitive data on premises without asking a small desktop to solve every computational problem.

96GB of memory changes what a mini PC can attempt

Local AI increasingly runs into memory limits before users exhaust processor features. Model weights, context, embeddings and intermediate data all need space. A machine with 16GB or 32GB can run useful small models, but larger models often require aggressive quantisation or offloading.

Acer’s 96GB ceiling gives the RI110 much more room. The company’s “up to 120B parameters” statement should still be read carefully. Parameter count alone does not establish performance. A 120-billion-parameter model at a low-bit quantisation may fit while running too slowly for interactive use, and different architectures have different memory overheads.

What the capacity does guarantee is flexibility. Developers can compare multiple model sizes, keep larger datasets resident and run more services simultaneously without treating memory as the first hard boundary.

Intel’s integrated architecture is doing several kinds of work

The Core Ultra X7 358H combines CPU cores, an NPU and Intel Arc B390 graphics. Those compute engines suit different workloads. The CPU handles general application logic, the NPU can accelerate efficient AI tasks, and the GPU provides highly parallel compute for graphics and many machine-learning operations.

Software decides whether that heterogeneity becomes useful or frustrating. Framework support, driver maturity and model runtimes need to place work on the right engine. A specification listing several accelerators does not mean every application automatically combines them.

Acer’s own Qubi Claw software is part of its attempt to make the machine easier to use for local agentic workloads. The more significant test will be compatibility with broader developer ecosystems, because businesses rarely want their AI strategy tied to one vendor utility.

OCuLink gives the small box an escape route

One of the RI110’s most interesting features is a 64Gbps OCuLink interface. OCuLink exposes PCI Express over a cable and can provide a lower-overhead path to external devices than interfaces designed primarily for general-purpose peripherals.

For a mini workstation, that creates the possibility of attaching external graphics, storage or specialised acceleration without putting full-size expansion slots inside the chassis. It does not make the RI110 as modular as a tower, but it narrows the gap.

External expansion has its own costs: another enclosure, power supply, cable and desk footprint. The feature is best seen as insurance. A user can start with the compact integrated system and add specialised hardware later if a workload outgrows the internal graphics.

Four terabytes of local storage supports private datasets

Up to 4TB of PCIe 4.0 storage is relevant because local AI is data-hungry. Model files can consume tens or hundreds of gigabytes, and organisations may need local copies of document collections, vector indexes, code repositories and generated media.

Keeping that data inside the workstation can reduce network dependence and create a clearer security boundary. It also transfers responsibility to the owner. Backups, encryption, access controls and retention policies become local operational tasks rather than services delegated to a cloud provider.

The same applies to model updates. A cloud API silently moves to new infrastructure; a local AI workstation needs its runtimes, model weights and security patches maintained deliberately.

Agentic AI makes reliability more important than benchmark peaks

An AI agent may run for minutes or hours, call tools, read files and wait on external events. That workload is different from a benchmark that pushes a processor at maximum throughput for a short test.

A machine used as an always-on agent host needs stable thermals, predictable sleep behaviour, networking and robust process supervision. Acer’s Veriton branding is significant because the family is aimed at business desktops rather than experimental enthusiast systems. Windows 11 Pro also provides enterprise management features absent from many hobbyist AI boxes.

Still, the security model deserves scrutiny. An agent with permission to read documents, send messages or execute code can amplify mistakes. Running it locally protects data from one set of external exposures but does not make unsafe automation safe. Organisations need least-privilege permissions, logging and human approval for consequential actions.

Hybrid computing is likely the realistic model

A local workstation is attractive for private data and predictable inference costs. Cloud AI remains attractive for models too large to fit locally, bursty workloads and services that need global scale. The RI110’s value lies in allowing a company to choose between those locations per task.

That choice can be particularly relevant where network latency, data-residency policies or bandwidth costs matter. South African organisations may find local processing useful even while relying on cloud regions for other applications. However, there is no verified South African pricing or availability for the RI110 yet, so the economics remain unknown.

The Veriton RI110’s most important specification is therefore not a TOPS figure. It is the combination of large memory, substantial local storage and an OCuLink expansion path inside a compact business machine. Those features make it plausible as infrastructure rather than merely a demonstration platform.

OCuLink is especially notable because it gives the RI110 an escape route from the limits of integrated graphics. The PCIe 4.0 x4 link can provide up to 64Gbps to an external GPU or high-speed storage device, avoiding some of the protocol overhead associated with more general-purpose external interfaces. That does not turn every eGPU into an internal graphics card, but it makes the mini workstation easier to adapt when a workload outgrows the Arc B390. Acer plans EMEA availability in the first quarter of 2027, so deployment decisions still depend on final regional pricing and support.

The workstation format also makes acoustics relevant. A local AI service may sit on a desk and run for long periods, so fan behaviour under sustained inference can matter more than momentary peak noise. Acer’s launch information does not establish sustained thermals or sound levels; those should be part of any later evaluation alongside raw model throughput.