ASUS's AI SuperBuild software turns the NUC 16 Pro mini-PC into a self-contained AI workstation, combining an Arc B390 GPU delivering 122 INT8 TOPS with a 50 INT8 TOPS neural processing unit for a total of 172 TOPS of local acceleration on a machine that fits on a desk.

The software, which ASUS built with Intel, targets users and developers who want to experiment with agentic AI and local large language models with minimal setup overhead. William Wong, Senior Content Director at Electronic Design, tested the stack on the NUC 16 Pro and walked away with a working local AI environment.

What hardware is doing the work?

The NUC 16 Pro runs on Intel's Core Ultra X9 series 3 processor, a 16-core CPU split across three tiers:

• Four performance cores for heavy threads
• Eight efficient cores for background load
• Four low-power cores for idle and housekeeping tasks

The Arc B390 GPU carries the bulk of AI throughput at 122 INT8 TOPS, while the on-chip NPU adds 50 INT8 TOPS. The NPU is the more efficient of the two, and because it sips power, it can run models in the background without dominating the thermal budget of the small-form-factor system.

How does the software handle models?

AI SuperBuild provides a straightforward download path to free large language models such as Qwen3, which Intel has already validated on the platform. It then supplies a chat interface that runs those models entirely on the local machine — no cloud inference, no subscription, no API key.

The more interesting layer is agentic AI support. SuperBuild implements the Model Context Protocol (MCP), an agent interface that Wong describes as "essentially device drivers that LLMs can use to access other data and services." In practice, this splits into two classes of agents:

• Local-only MCPs manipulate data on the machine and never touch the cloud
• Cloud-connected MCPs pull in cloud data and may offload processing to cloud services

SuperBuild helps users configure the system to use agents from the MCP marketplace, giving the mini-PC a path from simple chatbot duty toward multi-step, tool-using workflows.

What about non-programmers?

SuperBuild is primarily aimed at people who want to test AI at a lower level. Wong points to alternatives aimed at different audiences. OpenClaw targets non-programmers and maintains its own MCP marketplace focused on agentic AI tasks, which "may be more useful to many." ZeroClaw targets even smaller platforms with a minimal runtime footprint.

He also notes a market shift: most other free or open-source alternatives have moved to subscription models. The major platforms — Claude, Perplexity, and ChatGPT — remain dominant, but getting them set up for local operation can be challenging.

Where are the limits?

Local execution has clear advantages. Data and processing stay on the machine. In his testing, Wong got applications and agents running locally for search and other chores, and because nothing depended on remote servers, those actions kept working when the internet was down — or on a fully isolated system.

The performance ceiling is the first caveat. Wong is blunt: while the Core Ultra X9 is a powerful chip, "it doesn't compare in performance to what's in the cloud, with multiple GPGPUs and AI accelerators capable of handling extremely large LLMs." The NUC 16 Pro can run mid-size models comfortably, but frontier-class LLMs remain out of reach for this class of hardware. Even so, many practical applications fit within the local budget, and SuperBuild gives developers a way to test what is doable before committing to a deployment.

Security is the second caution. Running software locally means local data may be made available to the models, and that information may flow through MCPs. When an MCP runs in the cloud, users need to be careful about what data they expose and how it will be used — a local front end does not guarantee an air-gapped pipeline end to end.

What's the takeaway?

For engineering teams evaluating edge AI, the NUC 16 Pro plus AI SuperBuild offers a validated, zero-subscription testbed for agentic workflows on hardware that costs a fraction of a GPU server. Wong's testing confirms that local search and agent chores run reliably offline; whether Intel's MCP marketplace grows enough agent coverage to pull developers away from cloud-first platforms will determine how far this approach scales.