UGREEN HomeAgent HA100 First Look: Local AI at Home

At first glance, the UGREEN HomeAgent HA100 looks like a compact NAS. It has removable hard-drive bays, a large cooling system, a 2.5GbE network connection and the familiar proportions of a small home server. That description is technically correct, but it undersells what UGREEN is trying to build. HomeAgent combines several pieces of hardware that would normally exist as separate products: network storage for files and photos, an NVR for security cameras, a local AI server for language and vision models, and a smart-home platform that can run its own applications and services.

The idea behind the product becomes clearer once the whole system is set up. Instead of treating the cloud as the default destination for everything happening inside the home, HomeAgent is designed around doing much of the work locally. Camera recordings stay on local storage, image and face recognition can be processed locally, and even the voice assistant can run a small language model directly on the HA100. Cloud models are still available if the user wants them, but they are presented as an option rather than the foundation of the system.

Our test package consists of the HA100 itself, the compact SynCare ID500 Pro indoor camera and the considerably larger SynCare OD600 Pro outdoor camera. The cameras are important to understanding the product because HomeAgent is not simply a NAS that happens to support surveillance. The storage, camera system and AI hardware have been designed as separate but connected parts of the same platform.

The first thing that stands out is the industrial design. The HA100 does not look like a conventional black NAS enclosure designed to disappear into a utility closet or server rack. Its rounded white chassis, glossy black display strip and softly finished surfaces make it look much more like an appliance intended to remain visible in a living space. The design is restrained rather than decorative, and the small front display gives the unit some personality without turning it into a piece of gaming hardware.

The screen first shows a simple UGREEN boot animation and, once the machine is running, becomes a status display with information such as the time and system status. A small green light is projected onto the surface beneath the chassis. None of these details is essential to the operation of the product, but together they help distinguish HomeAgent from the utilitarian appearance of most NAS hardware.

There is another storage location underneath the HA100. Opening the bottom panel exposes an M.2 2280 NVMe SSD slot, and UGREEN includes a thermal pad in the package for SSD installation.

This SSD appears to have a different role from the three large SATA bays. The hard drives are the obvious choice for bulk storage and continuous camera footage, while the faster NVMe slot is better suited to applications, AI models and other workloads that benefit from lower latency and higher random-access performance. That division starts to make more sense once the software is running because HomeAgent is not simply storing files; it is downloading multi-gigabyte models, running Docker-based services and processing local AI workloads.

I did not have an NVMe SSD available during this initial setup, so the HA100 was tested with hard drives only. The machine works without an SSD, and nothing in the initial setup prevented the system from running. Whether NVMe storage makes a meaningful difference to model loading, application responsiveness or AI performance is something we will test later rather than speculate about here.

The SynCare ID500 Pro and OD600 Pro follow very different design philosophies. The ID500 Pro is a compact indoor camera with a motorized pan-and-tilt mechanism. It is small enough to sit on a shelf or cabinet without drawing much attention, which matters for a device that may spend its entire life in a living room, hallway or home office.

The regulatory label on our unit specifies a 5V/2A input and clearly marks the product for indoor use. It also carries FCC, CE, UKCA and other regional compliance markings. The package includes the camera, power supply, cable and mounting accessories, so it can either be placed on a surface or installed more permanently.

The OD600 Pro is substantially larger and looks much more like a dedicated surveillance product. Its multi-camera front assembly and larger housing immediately set it apart from the small indoor unit, and the enclosure is marked IP66 for outdoor use.

The physical setup was straightforward. The software took more patience.

UGREEN supplied the mobile application through Apple's TestFlight, and I initially installed the latest version which is 1.20.20. During the activation process, after Model Manager page, the app reached the Voice Assistant setup screen and repeatedly returned a “Parameter error!” message while click on Continue button. The error prevented activation from continuing, so this was not simply a cosmetic problem inside an otherwise functional interface.

Once activation finally started, the process took close to half an hour. My initial assumption was that the HA100 was downloading its large AI models during this time, but that turned out not to be the case. After logging into the finished system, Model Manager still required the main language model to be downloaded separately.

The current build offers Qwen3.5 4B as the local language model, with a listed download size of 4.3GB. The setup screen also lists separate voice and intent-recognition components, which means the language model is only one part of the local AI stack.

That makes the long activation time somewhat difficult to explain from the user's perspective. The HA100 is clearly configuring a relatively complicated software environment in the background, but the app does not currently provide much detail about what is happening. A more descriptive progress screen would make a 20- or 30-minute initialization period feel much less mysterious.

After activation, additional components still had to be installed. Rather than monitoring every download, I left the HA100 running overnight. By the following morning, the necessary services had finished installing and the system was ready for the camera to be connected. The automation was actually quite good once the initial activation problem was behind us: there are many moving parts underneath HomeAgent, but most of them were deployed without requiring constant user intervention.

The Model Management interface gives a much clearer picture of what UGREEN means by “local AI.” HomeAgent does not rely on one general-purpose chatbot for every task. Instead, the software exposes separate categories for large language models, people recognition, text recognition, similar or duplicate photo detection, image recognition, voice chat, intent recognition, face recognition and general detection.

That architecture makes sense for an edge device. A specialized computer-vision model does not need the overhead of a general-purpose LLM just to identify a person, vehicle or object, while a language model is useful when the user wants to interact with the system conversationally. HomeAgent appears to combine several specialized models and then use the assistant layer to connect them.

The current local language model page is particularly interesting. The app lists Qwen3.5 4B, a 4.3GB model, with approximately 2GB of memory usage and a claimed response speed of 40 tokens per second.

Those numbers are values displayed by UGREEN's software, not results from our own benchmark, so they should not yet be treated as measured performance. We will test inference speed separately in the follow-up article. What they do show is that HomeAgent is genuinely running a local language model rather than simply presenting a cloud chatbot through a smart-home interface.

Interestingly, UGREEN does not treat local and cloud AI as mutually exclusive. The software also includes a cloud-model configuration page where an OpenAI API key and API endpoint can be entered manually.

That is probably the more practical architecture. Some tasks make obvious sense locally: face recognition, basic object detection, camera-event analysis and simple home-control commands benefit from low latency and from keeping sensitive data inside the home. More demanding reasoning or generative tasks may still be better suited to a larger cloud model.

The important point is that HomeAgent gives the user both options. Local processing is part of the architecture rather than a marketing label, but the platform does not artificially prevent cloud integration when it may be useful.

How smoothly the two approaches work together is something we still need to test.

HomeAgent becomes even more interesting inside the App Center. Instead of presenting the HA100 as one fixed application, UGREEN exposes a collection of installable components, including Uliya, Surveillance Center, Smart Home, Sync & Backup, Theater, TextEdit, Task Manager and the T20 AI Accelerator Card Toolkit.

The Smart Home application is still labeled Beta, and its information page explicitly lists Docker as an associated application. Docker also appears separately in the system interface, which strongly suggests that part of the HomeAgent software stack is containerized.

For technically inclined users, this is one of the more intriguing aspects of the HA100. The product behaves less like a fixed smart-home appliance and more like a small home server with a consumer interface layered over it. Whether UGREEN eventually opens more of this environment to third-party applications will have a major effect on how flexible HomeAgent becomes, although that possibility should be treated as potential rather than something the company has promised.

Once the required services were installed, the ID500 Pro appeared correctly inside Surveillance Center and could be managed through the HomeAgent interface. That part of the setup was considerably smoother than the initial activation. The camera and HA100 clearly feel like parts of the same ecosystem rather than two unrelated products connected through a generic protocol.

The HomeAgent software also includes a voice assistant called Uliya. In the current build, the setup screen offers English, German and Mandarin, and UGREEN presents examples that include home control, security-footage search and music playback.

The security-search example is especially revealing because a request such as “Was anyone at the door yesterday at 5 PM?” requires several parts of the system to work together. The assistant has to understand the question, identify that it refers to surveillance footage, search locally stored video or event metadata and return a useful result. That is a much more interesting use of local AI than simply adding a chatbot to a NAS.

Whether these interactions work consistently will be one of the most important questions for the next stage of testing. The hardware architecture already makes sense; the challenge now is whether the software can turn all of those individual capabilities into something that actually feels simpler than using several separate devices and apps.

The most significant issue so far is the activation failure in app version 1.20.20. A bug that blocks onboarding entirely matters much more than a mislabeled button or an unfinished animation. Rolling back to 1.19.20 solved the problem in my case, but version management should obviously not be part of the normal setup process once the product reaches consumers.

The second issue is the amount of time required for the initial configuration. Activation took nearly half an hour even though the 4.3GB Qwen model still needed to be downloaded afterward. HomeAgent is clearly doing a substantial amount of setup behind the scenes, but the current interface does not explain enough about that process.

Fan noise is another area I will continue to watch. At idle, the HA100 is relatively quiet. Under heavier activity, however, the cooling fan becomes clearly audible. Interestingly, the sound is more acceptable with the magnetic top cover removed; replacing the cover seems to amplify or redirect some of the noise. I would not currently place the HA100 close to a bed if it is expected to perform downloads, indexing or AI workloads overnight.

The OD600 Pro also presents a different kind of concern. Its wired PoE design makes technical sense for permanent surveillance, but the installation process may be a barrier for the U.S. consumer market, where battery and solar cameras have trained users to expect an outdoor camera to be installed with relatively little wiring. This is more a question of product positioning than a hardware flaw, but it is an important one.

After unboxing and setting up the HomeAgent system, the clearest impression is that the hardware feels more mature than the software.

The physical product is unusually well thought out. The magnetic panels, accessible drive trays, color-coded storage architecture, removable AI accelerator, dedicated surveillance disk, M.2 expansion and integrated PoE connectivity all feel like parts of a coherent design rather than features added to fill out a specification sheet. Even small details such as the drive locks, SSD thermal pad and blank cover for the accelerator slot suggest that UGREEN has spent real time thinking about installation and maintenance.

The software is more ambitious than I expected, but it is also where the unfinished parts are most visible. The activation bug in TestFlight 1.20.20 is significant, several applications remain in beta, and some parts of the onboarding experience expose the complexity underneath the platform more than a consumer product ideally should.

At the same time, that underlying complexity is exactly what makes HomeAgent interesting. The system already exposes a local language model, multiple specialized recognition models, Docker-based applications, a dedicated AI accelerator, local surveillance storage and optional cloud-model integration. UGREEN is effectively trying to turn a small home server into something that can be used like a consumer appliance.

The more difficult question now is not whether UGREEN can build the hardware. The HA100 already makes a convincing case there. The next question is whether the software can make all of that hardware genuinely useful: whether local AI can search camera footage accurately, whether the assistant can move naturally between smart-home control and stored information, how much the accelerator changes performance, and whether the system remains easy to live with after the initial novelty wears off.

Those are the questions for the next article.