Local LLMs on Synology DS224+: Is it Even Possible?

Running local Large Language Models (LLMs) on a Synology DS224+ NAS is a question that keeps coming up in the self-hosting community. The short answer is: it’s possible, but you need to manage expectations. This isn’t about running the latest, biggest models; it’s about understanding the hardware limitations and what you can realistically achieve for home AI use.

We tested the limits of the DS224+ to see what kind of local LLM experience is actually viable on a 2-bay NAS. The core challenge here isn’t the software—the tools for running these models are getting better—it’s the hardware itself. The DS224+ is designed to be reliable storage, not a dedicated AI compute cluster.

The Hardware Bottleneck: CPU vs. GPU

When you talk about running LLMs, you are talking about massive amounts of matrix multiplication. This process demands serious compute power. Most consumer-grade AI acceleration comes from dedicated GPUs. The DS224+ relies on its built-in CPU and whatever RAM it has access to. This immediately sets a hard ceiling on model size and inference speed.

If you want to run anything substantial, you are going to be fighting the CPU clock speed and the available RAM. Trying to load a large, state-of-the-art model will likely result in painfully slow token generation—we’re talking minutes per answer, not seconds.

What Actually Runs? The Tradeoff

Our testing showed that success hinges on model quantization and size. You aren’t going to run a 70-billion parameter model. Instead, the sweet spot seems to be much smaller, highly quantized models. These are the models that have been aggressively compressed to run on less powerful hardware.

For practical, day-to-day use—like summarizing a document you just downloaded or answering basic questions based on local files—a smaller, optimized model running via a containerized setup is where you’ll find the most usable results. The tradeoff is obvious: smaller model size means less complex reasoning capability.

Setting Up the Environment

Getting this running requires more than just installing an app. You are dealing with containerization, likely Docker or a similar environment, which adds layers of complexity. The goal is to get the necessary inference engine running efficiently, minimizing overhead so that the CPU cycles are dedicated to the math, not to the operating system fighting for resources.

For the home user who just wants to type a prompt and get a quick answer, the experience on the DS224+ is going to feel sluggish. It’s a proof-of-concept machine, not a production AI workstation. If your use case requires near-instantaneous, complex reasoning, you’ll need to look at dedicated mini-PCs or systems with discrete, powerful GPUs. The NAS is best viewed as the secure, always-on storage layer for your data, with the LLM running on it being a highly constrained, educational experiment.