Zettlab D6 Ultra: 30-second review

The world of NAS systems has been turned on its head in the past couple of years, with Ugreen entering the fray aggressively and Synology walking away from the prosumer market.

In what appears to be an attempt to rekindle the same spark that propelled Ugreen, ZettLabs is launching a range of NAS on Kickstarter, with a focus on personal AI.

Having already launched two ARM-based NAS, the D4 and D6, the two new AI models use Intel processors and are the six-bay D6 Ultra reviewer here, and an eight-bay D8 Ultra. These are both available through a pre-order system and can be purchased either barebones with no memory or with DDR5 pre-installed.

At the heart of the D6 Ultra is the Intel Core Ultra 5 Processor 125H, a series 100 processor that first appeared in laptops back in late 2023.

This platform is powerful enough to handle the six conventional drive bays, the dual 10GbE LAN ports, and the dual USB4 ports that this NAS offers. The memory installed model comes with 32GB of DDR5, but this can be upgraded to 96GB for those who can afford 48GB modules.

Where this diverges from other six-bay NAS is that, with the Intel Core Ultra 5 Processor 125H and its Intel AI Boost capabilities, this system can host LLM AI models and run them in isolation. AI is a niche requirement, but those who don’t use AI are still going to get a fast, powerful NAS to share files, make backups, and interact with Cloud services.

With this level of system-resident functionality in a NAS, the price is higher than that of a conventional 6-bay NAS, so it will only be of interest to those who want its AI capabilities. However, Zettlabs aren’t the only NAS maker offering the hosting of local AI models, and the general features of ZettOS aren’t at the same level as more established solutions.

I wouldn’t write off the Zettlab D6 Ultra as one of our best NAS devices in years to come, but the operating system needs to mature before that happens.

Zettlab D6 Ultra: Price and availability

• How much does it cost? From $1680
• When is it out? On pre-order
• Where can you get it? Direct from Zettlab

This Zettlab D6 Ultra is currently on pre-order from the company's website and comes in two variants, with and without memory. The version I tested for this review came with 32GB of DDR5 memory and costs $1679.99, and the same hardware without that RAM is only $1079.99.

That’s either a reflection of how much RAM costs these days, or how much Zettlab is willing to charge you for it. As I was able to find Crucial 16GB SODIMMs for around $150 on Amazon.com, taking the thirty seconds to populate this machine yourself could easily save you $300.

For those interested, the 8-bay D8 Ultra, is priced at $1319.99 with no memory, and $1919.99 for 32GB, which is a similar price differential for two memory modules.

What might be more problematic for Zettlab is that the Super Early Bird pricing of the Ugreen NASync iDX60011 Pro with 64GB of DDR5 is only $1559, while the MSRP is $2599.

The iDX60011 Pro is also a 6-bay NAS, built on Intel mobile silicon, but it uses the more powerful Core Ultra 7 255H, a modern 200-series processor.

I haven’t tried that NAS yet, but it arrived today, so soon I should have a baseline for comparing the two platforms. But on paper, the D6 Ultra does seem expensive when supplied with RAM, and the iDX60011 Pro has a potential performance advantage.

• Value: 3.5 / 5

Zettlab D6 Ultra: Specs

Zettlab D6 Ultra: Design

• Solid construction
• Great port selection
• Minor tweaks needed

The design and layout of the D6 Ultra aren’t radical, but the chassis's all-metal construction suggests this is a machine with a long life ahead.

However, if this NAS had been entirely metal, it would be exceptionally heavy, and the six drive trays are made of plastic.

Curiously, the trays are labelled A through F and are not numbered. What’s nice about the tray design is that for 3.5-inch drive installations, no tools or screws are required. What I didn’t care for is that they don’t include any sort of locking mechanism, and triggering them to open requires only a light press.

Given the utter chaos that disconnected drives can cause in a running system, these drive trays need locks or a mechanism to prevent all trays from opening accidentally.

A feature of all Zettlab NAS is the 3.49-inch display at the bottom left of the fascia, which shows drive status, network IP address, and more. Most people will need to get reasonably close to read the information from this display, but it’s a good alternative to flashing LEDs.

Also on the front are two card readers, one is SD4.0 and the other TF4.0, covering both common card types. And alongside those are a USB 3.2 Gen 2 port and a USB4 Type-C port. A feature that initially confused me was the button on the far right of the fascia, which I easily assumed powered the NAS up. It doesn’t.

The power button is on the back, out of the way. The button on the front is designed to initiate copying files from SD cards and USB ports to the internal storage.

On the rear are another USB 3.2 Gen 2 port, another USB4 port, a USB 2.0 port, dual 10GbE LAN ports, and an HDMI out. These are all to be expected on a NAS at this price point, but what I’ve not seen before is the SFF-8654, a port built to provide an external connection for PCIE4.0 card with 8 lanes.

Using that port, it should be possible to connect an external GPU, expand the storage, or install a 25GbE network adapter. While there isn’t the physical room inside the NAS for a full-sized video card, the SFF-8654 enables one to be outside with enough bandwidth to the system to be useful.

Based purely on the included ports, the specification of this machine was carefully designed to please those who use NAS systems to their fullest potential.

My only concern about the hardware is that a few minor details suggest the D6 Ultra changed extensively during development, and the industrial engineers involved struggled to keep up with those modifications.

An obvious mistake I noticed was that the magnetically attached filter that covers the fans on the back doesn’t fit correctly. It’s too small, and slides down when attached.

Another is that, underneath the NAS, there is an access panel that provides access to the two M.2 and two SODIMM memory slots. The plate that covers this has four screws retaining it, when one or two screws would have been sufficient.

That’s a minor thing, but what’s more of an issue is that Zettlab provides two thermal pads to place on M.2 drives to connect them thermally to the skin of the D6 Ultra. Unfortunately, these pads are far too thick, and if four screws are tightened down, they could put excessive pressure on the NVMe drives to the mainboard, causing damage.

This configuration also doesn’t account for NVMe SSDs that have a heatsink attached.

I hope the filter and the thermal pads both get addressed when the D6 Ultra next has a version change, because the cost of this NAS dictates that the details are right.

• Design: 3.5 / 5

Zettlab D6 Ultra: Features

• Intel Core Ultra 5 125H
• 28 PCIe Lanes
• Intel Deep Learning

For many years, NAS makers almost exclusively used either ARM SoCs or, occasionally, low-power Intel chips like the Atom or Celeron series.

The design logic for this was sound, since moving data from SATA drives doesn’t require much computing power.

What’s happened more recently is that the app installations on NAS have become much more sophisticated, with Virtual Machine and Docker containers being used, but also now we’re transitioning into an era where NAS are AI nodes curating the data they hold.

As a result, we’re seeing more machines like the D6 Ultra, which use repurposed mobile platforms like the Intel Core Ultra 5 125H, a processor with 14 cores, a 7 Xe core GPU, and dedicated AI silicon.

This is far from the most powerful CPU that I’ve seen in a NAS, but the functionality that it inherently comes with because of this Meteor Lake generation processor casts a long shadow over those NAS designs still relying on ARM CPUs or Intel N300 chips.

The AI component in the Core Ultra is an important aspect that elevates it above lesser Intel silicon and ARM SoCs, but the feature of this hardware that has a greater impact, I’d suggest, is the 28 PCIe lanes.

The Zettlan support documentation helpfully reveals how those lanes are allocated, taking the guesswork out of how the bandwidth pie is sliced up.

For starters, each of the two M.2 SSDs is configured as PCIE 4.0 x4, which could make them capable of up to 7,000Mbps transfers if used directly as storage. That’s eight lanes, and a further four are given over to the 10GbE LAN ports, with two lanes per port.

That’s twelve used up, another eight are allocated to the SFF-8654, and two are used for the SATA interface that the hard drives attach to. That leaves two lanes for the card readers and other minor requirements. It’s my understanding that the USB ports are all inherent to the CPU, so they don’t need PCIe lanes, but I could be wrong about that.

If all that is accurate, then this is one of the few NAS I’ve tested where most of the PCIe bandwidth is utilised, on a platform that has plenty to hand out.

However, this technical achievement isn’t the focus of the Zettlab marketing, because the favourite buzzword of the moment is AI, and the Core Ultra 5 does bring reasonable offerings to the AI table.

Where the D4 and D6 models have ARM processors with 6 TOPS (Trillions/Tera Operations Per Second) of AI processing, the D6 Ultra and its Intel Core Ultra 5 125H have 34 TOPS. That number is a combination of the CPU, GPU, and NPU, with Intel’s AI Boost silicon contributing 11 TOPS to the total.

That’s significantly better than the ARM chips, although compared with the likes of the Nvidia high-end GPUs, like the RX 5090, which can muster 3,352 TOPS, it's still at the modest end of the scale. However, this NAS is sufficient for running local AI models, and using the SFF-8654 port, external GPUs can be added to significantly bolster AI capability.

Overall, the hardware in the D6 Ultra is impressive, even if Intel has released better chips since the Meteor Lake era.

• Features: 3.5 / 5

Zettlab D6 Ultra: Software

• ZettOS
• ZettAI

Since the likes of Synology, Qnap and Asustor all have mature NAS operating systems, the best comparisons can be made between ZettOS and Ugreen’s evolving UGOS Pro operating system.

My immediate reaction to ZettOS was that even in this early stage, it has features that took at least six months or longer to appear on UGOS Pro, and a much better app selection.

These include support for Docker and Virtual Machines, media tools, Home Assistant, Plex, Jellyfin, Unifi, and a collection of developer tools.

While I’d have expected to see more software development options, the inclusion of Docket and VM provides an easy means to add those things either with a container or a VM of a desktop Linux distro.

I’m not a huge fan of the red, yellow, green dots for window controls, aping Apple; the Web interface is relatively clean and doesn’t require supporting documentation to navigate.

But there are a few significant holes in the feature selection of the OS, most notably with respect to security. At the time of writing, there is no 2FA, limiting access to the machine via a login and password, and if you use the Windows Zettlab AI NAS app, those are both stored on the client PC. It is possible to use a Zettlan Remote ID to connect to the NAS externally, using the Zettlab cloud portal.

I was a little shocked by some of the security choices made for this unit when I discovered that by default, the FTP server functionality was active. That’s not typically considered a wise move, and admins only activate that feature when they’ve put in place controls to avoid it being externally exploited.

Another area where this NAS OS veers slightly off the beaten path is its file system, which, to my understanding, is a proprietary one developed by Zettlab. Those expecting the choice of Ext4 or BTRFS will be disappointed, and I don’t think the current file system supports a hybrid structure with drives of different sizes. For sharing, SMB and NFS are supported, but I didn’t see any means to format USB-connected storage. In fact, all the external drives I connected, either to USB 3.2 or USB4 ports, were ignored. Eventually, I got a thumb drive that appeared to be formatted in FAT32, but drives that were preformatted in exFAT or NTFS were not recognised. That’s a feature that needs to be made a priority, I’d suggest.

Having two 10GbE LAN ports offers some great network bandwidth, but there currently aren’t any link aggregation or failover options to leverage the full potential of them.

In my tests, the USB4 ports did not work in host mode, although this feature, I believe, is promised. The USB-C ports did charge my laptop, at least.

The HDMI port does nothing currently, not even showing the Linux boot.

During my time with this machine, it underwent two firmware updates, suggesting that the software developers are backfilling functionality that’s either missing or not working optimally.

• Software: 3.5/5

Zettlab D6 Ultra: Performance

• Network performance
• External drives limitation
• AI models

For my testing, I used six IronWolf 4TB drives and allocated them into a RAID 10 pack for the best possible speed available. And to enhance that further, I allocated one Crucial P3 NVMe drive as a 1TB cache.

Like with many NAS, for whatever reason, a single drive can only be allocated to caching reads, and it takes two modules to cache reading and writing. I didn’t have two spare M.2 drives, so I went with the cached reading instead.

Over a single 10GbE, the read throughput hit over 900MB/s, which is excellent. As there is no link aggregation on the network ports, that’s realistically as fast as it’s possible to go. Should host mode on the USB4 be made active, that should be capable of much faster speeds. But without caching, six hard drives hit a bandwidth ceiling of around 900MB/s, since each is only capable of about 150MB/s. For this reason, unless you run SATA SSDs or have large M.2 cache drives for both reading and writing, there is little point in using the SFF-8654 port to add more or faster network ports.

And, in the support material, it states that “We currently do not natively support U.2 or U.3, but our machines can expand via the SFF-8654 interface.” Since SFF-8654’s function on some motherboards is to connect U.2 or U.3, that seems an odd choice.

As I mentioned earlier, external drives' functionality is incomplete, and without support for NTFS and exFAT, it's extremely limited in what it can be used for. I also found it disappointing that if the system didn’t recognise the file format, it didn’t offer to format it into one it was happy to work with.

That fun was as if nothing compared to the adventure of using this NAS as an AI local platform.

Inherently, the D6 Ultra and D8 Ultra support a local LLM model that can analyse whatever documents you put on the NAS, providing a chatbot interface to work with the contents.

If that sounds great to you and you have lots of files you need to navigate with AI logic, then right out of the box, this might be the NAS for you.

When the processing of the local model occurs is configurable, so that it doesn’t step on current tasks like file serving.

What’s great about this functionality is how automatic some of it is. After loading some prior review content folders onto the NAS, I discovered that the LLM had created a photo album based on the files and the subjects that it saw in the images. It could then also answer questions about the files, revealing the knowledge it has gained processing them.

Or rather, that’s what is implied. Except when you ask it in the AI chat window, it wants you to specifically say which files it should check, which isn’t super-helpful.

I should also say that the default ZettAI created by Zettlab, I assume, was poor at some general AI tasks, like history.

To further explore this feature, I looked at all the models that the system has available to install. These included four variants of Gemma, the Google AI, four flavours of Phi, the open-source Small Language Models (SLMs) created by Microsoft, two more QwQ models made by the Owen Team, two DeepSeek-R1 models, and a couple of Meta-made Llama models.

It’s possible to load and use each of these, though they range in size from about 2GB to more than 4GB, and some use plenty of memory.

I tried a number of them, and to put it mildly, my mileage varied considerably.

What I learned was to not ask DeepSeek about history, since it failed the most basic questions about historical events.

Its collection of the Kings and Queens of England was horrifically wrong, with it deciding that Queen Elizabeth I reigned from 1237 to 1558, which would have easily made her the oldest person ever recorded, if it were true. For those wondering, her reign lasted from 1558 to 1603.

Realizing what a rich mine of alternative information DeepSeek could be, I then asked about which US presidents died in office, and it completely messed up that challenge. It said ten presidents had died while in office, whereas the right answer is eight. It got the names of those eight wrong, included people who died after they left office, and insisted that three presidents died while hunting.

ZettaAI did a better job of the King's question, although not perfect, as it left out Harold, who died at Hastings. But it entirely messed up the dead Presidents, leaving out Zachary Taylor, Abraham Lincoln, James A. Garfield, Warren G. Harding and even John F. Kennedy. Then it included Andrew Jackson, Martin Van Buren, Theodore Roosevelt, Woodrow Wilson, Dwight D. Eisenhower, Lyndon B. Johnson, Richard Nixon and Gerald Ford, none of whom died in office. Out of eleven U.S. presidents it provided, the only ones it got right were William Henry Harrison and William McKinley.

I’m sure it’s possible to find subjects that these AIs are much better with than Western history, but the point is that if you were a student using these tools for homework exercises, you could be in deep, deep trouble.

To be clear, the effectiveness of these models, or not, isn’t a reflection on the Zettlab D6 Ultra, but the nature of AI technology, and its value to those expecting it to come back with generally correct answers.

When I questioned the ZettAI about how Abraham Lincoln wasn’t in the list of Presidents in one of those who were assassinated, it tried to say that the list of those who died in office didn’t include those who were killed, even though that was not a context I created. Then contradicting itself, it also argued that Lincoln died of Pneumonia, caused by the gunshot, but not directly from the assassination. This appears to be a riff on the concept that guns don’t kill people; complications from gunshot wounds kill people.

In short, if you are expecting something as powerful as datacentre AI in a small box on your desktop, you might need to scale that objective back that thinking somewhat, though as models improve, it might become an invaluable tool.

What I also need to say is that most free AI tools are significantly behind the curve in terms of hallucinations and other issues. It's hard to say if models like ChatGPT 5.1 and Claude 4 will come to this platform, but it would make this experience dramatically better if they did.

• Performance: 4 / 5

Zettlab D6 Ultra: Final verdict

There are some positive things to say about the D6 Ultra, since, for a NAS platform in the earlier stages of development, ZettOS is already reasonably sophisticated.

Where more questions exist is in the value of AI on hardware like this, because, as nice as a Core Ultra 5 CPU is, it's hardly a data centre. Depending on the size of the dataset you wish to use with the AI, this could be a highly responsive and productive solution, or something painfully slow to access.

I’m aware that even more powerful NASs are coming along that can outperform the D6 Ultra, although if an external GPU were added to this platform, it might be quicker. What’s not a guess is that if you added an RTX 5090 to this machine externally and gave it 96GB of RAM to run its models, this could be an impressive local AI solution, but the system's cost would be a minor part of the total expenditure.

Given the power of datacentre AI solutions, a solution like the D6 Ultra is likely to interest only those who want to use models experimentally or isolate the development of an AI platform from the Internet.

But I have to question how cost-effective this would be in the long term, should the model reach a level of complexity where the NAS struggles to run it interactively to achieve the level of performance you might want.

Should the AI bubble burst and people realise that it's of limited use for many tasks, at least this hardware is sufficient to be an excellent file server and media system.

Should you buy a Zettlab D6 Ultra?

Buy it if...

Don't buy it if...

For more NAS solutions, these are the best NAS hard drives we've reviewed.

Zettlab D6 Ultra: 30-second review

The world of NAS systems has been turned on its head in the past couple of years, with Ugreen entering the fray aggressively and Synology walking away from the prosumer market.

In what appears to be an attempt to rekindle the same spark that propelled Ugreen, ZettLabs is launching a range of NAS on Kickstarter, with a focus on personal AI.

Having already launched two ARM-based NAS, the D4 and D6, the two new AI models use Intel processors and are the six-bay D6 Ultra reviewer here, and an eight-bay D8 Ultra. These are both available through a pre-order system and can be purchased either barebones with no memory or with DDR5 pre-installed.

At the heart of the D6 Ultra is the Intel Core Ultra 5 Processor 125H, a series 100 processor that first appeared in laptops back in late 2023.

This platform is powerful enough to handle the six conventional drive bays, the dual 10GbE LAN ports, and the dual USB4 ports that this NAS offers. The memory installed model comes with 32GB of DDR5, but this can be upgraded to 96GB for those who can afford 48GB modules.

Where this diverges from other six-bay NAS is that, with the Intel Core Ultra 5 Processor 125H and its Intel AI Boost capabilities, this system can host LLM AI models and run them in isolation. AI is a niche requirement, but those who don’t use AI are still going to get a fast, powerful NAS to share files, make backups, and interact with Cloud services.

With this level of system-resident functionality in a NAS, the price is higher than that of a conventional 6-bay NAS, so it will only be of interest to those who want its AI capabilities. However, Zettlabs aren’t the only NAS maker offering the hosting of local AI models, and the general features of ZettOS aren’t at the same level as more established solutions.

I wouldn’t write off the Zettlab D6 Ultra as one of our best NAS devices in years to come, but the operating system needs to mature before that happens.

Zettlab D6 Ultra: Price and availability

• How much does it cost? From $1680
• When is it out? On pre-order
• Where can you get it? Direct from Zettlab

This Zettlab D6 Ultra is currently on pre-order from the company's website and comes in two variants, with and without memory. The version I tested for this review came with 32GB of DDR5 memory and costs $1679.99, and the same hardware without that RAM is only $1079.99.

That’s either a reflection of how much RAM costs these days, or how much Zettlab is willing to charge you for it. As I was able to find Crucial 16GB SODIMMs for around $150 on Amazon.com, taking the thirty seconds to populate this machine yourself could easily save you $300.

For those interested, the 8-bay D8 Ultra, is priced at $1319.99 with no memory, and $1919.99 for 32GB, which is a similar price differential for two memory modules.

What might be more problematic for Zettlab is that the Super Early Bird pricing of the Ugreen NASync iDX60011 Pro with 64GB of DDR5 is only $1559, while the MSRP is $2599.

The iDX60011 Pro is also a 6-bay NAS, built on Intel mobile silicon, but it uses the more powerful Core Ultra 7 255H, a modern 200-series processor.

I haven’t tried that NAS yet, but it arrived today, so soon I should have a baseline for comparing the two platforms. But on paper, the D6 Ultra does seem expensive when supplied with RAM, and the iDX60011 Pro has a potential performance advantage.

• Value: 3.5 / 5

Zettlab D6 Ultra: Specs

Zettlab D6 Ultra: Design

• Solid construction
• Great port selection
• Minor tweaks needed

The design and layout of the D6 Ultra aren’t radical, but the chassis's all-metal construction suggests this is a machine with a long life ahead.

However, if this NAS had been entirely metal, it would be exceptionally heavy, and the six drive trays are made of plastic.

Curiously, the trays are labelled A through F and are not numbered. What’s nice about the tray design is that for 3.5-inch drive installations, no tools or screws are required. What I didn’t care for is that they don’t include any sort of locking mechanism, and triggering them to open requires only a light press.

Given the utter chaos that disconnected drives can cause in a running system, these drive trays need locks or a mechanism to prevent all trays from opening accidentally.

A feature of all Zettlab NAS is the 3.49-inch display at the bottom left of the fascia, which shows drive status, network IP address, and more. Most people will need to get reasonably close to read the information from this display, but it’s a good alternative to flashing LEDs.

Also on the front are two card readers, one is SD4.0 and the other TF4.0, covering both common card types. And alongside those are a USB 3.2 Gen 2 port and a USB4 Type-C port. A feature that initially confused me was the button on the far right of the fascia, which I easily assumed powered the NAS up. It doesn’t.

The power button is on the back, out of the way. The button on the front is designed to initiate copying files from SD cards and USB ports to the internal storage.

On the rear are another USB 3.2 Gen 2 port, another USB4 port, a USB 2.0 port, dual 10GbE LAN ports, and an HDMI out. These are all to be expected on a NAS at this price point, but what I’ve not seen before is the SFF-8654, a port built to provide an external connection for PCIE4.0 card with 8 lanes.

Using that port, it should be possible to connect an external GPU, expand the storage, or install a 25GbE network adapter. While there isn’t the physical room inside the NAS for a full-sized video card, the SFF-8654 enables one to be outside with enough bandwidth to the system to be useful.

Based purely on the included ports, the specification of this machine was carefully designed to please those who use NAS systems to their fullest potential.

My only concern about the hardware is that a few minor details suggest the D6 Ultra changed extensively during development, and the industrial engineers involved struggled to keep up with those modifications.

An obvious mistake I noticed was that the magnetically attached filter that covers the fans on the back doesn’t fit correctly. It’s too small, and slides down when attached.

Another is that, underneath the NAS, there is an access panel that provides access to the two M.2 and two SODIMM memory slots. The plate that covers this has four screws retaining it, when one or two screws would have been sufficient.

That’s a minor thing, but what’s more of an issue is that Zettlab provides two thermal pads to place on M.2 drives to connect them thermally to the skin of the D6 Ultra. Unfortunately, these pads are far too thick, and if four screws are tightened down, they could put excessive pressure on the NVMe drives to the mainboard, causing damage.

This configuration also doesn’t account for NVMe SSDs that have a heatsink attached.

I hope the filter and the thermal pads both get addressed when the D6 Ultra next has a version change, because the cost of this NAS dictates that the details are right.

• Design: 3.5 / 5

Zettlab D6 Ultra: Features

• Intel Core Ultra 5 125H
• 28 PCIe Lanes
• Intel Deep Learning

For many years, NAS makers almost exclusively used either ARM SoCs or, occasionally, low-power Intel chips like the Atom or Celeron series.

The design logic for this was sound, since moving data from SATA drives doesn’t require much computing power.

What’s happened more recently is that the app installations on NAS have become much more sophisticated, with Virtual Machine and Docker containers being used, but also now we’re transitioning into an era where NAS are AI nodes curating the data they hold.

As a result, we’re seeing more machines like the D6 Ultra, which use repurposed mobile platforms like the Intel Core Ultra 5 125H, a processor with 14 cores, a 7 Xe core GPU, and dedicated AI silicon.

This is far from the most powerful CPU that I’ve seen in a NAS, but the functionality that it inherently comes with because of this Meteor Lake generation processor casts a long shadow over those NAS designs still relying on ARM CPUs or Intel N300 chips.

The AI component in the Core Ultra is an important aspect that elevates it above lesser Intel silicon and ARM SoCs, but the feature of this hardware that has a greater impact, I’d suggest, is the 28 PCIe lanes.

The Zettlan support documentation helpfully reveals how those lanes are allocated, taking the guesswork out of how the bandwidth pie is sliced up.

For starters, each of the two M.2 SSDs is configured as PCIE 4.0 x4, which could make them capable of up to 7,000Mbps transfers if used directly as storage. That’s eight lanes, and a further four are given over to the 10GbE LAN ports, with two lanes per port.

That’s twelve used up, another eight are allocated to the SFF-8654, and two are used for the SATA interface that the hard drives attach to. That leaves two lanes for the card readers and other minor requirements. It’s my understanding that the USB ports are all inherent to the CPU, so they don’t need PCIe lanes, but I could be wrong about that.

If all that is accurate, then this is one of the few NAS I’ve tested where most of the PCIe bandwidth is utilised, on a platform that has plenty to hand out.

However, this technical achievement isn’t the focus of the Zettlab marketing, because the favourite buzzword of the moment is AI, and the Core Ultra 5 does bring reasonable offerings to the AI table.

Where the D4 and D6 models have ARM processors with 6 TOPS (Trillions/Tera Operations Per Second) of AI processing, the D6 Ultra and its Intel Core Ultra 5 125H have 34 TOPS. That number is a combination of the CPU, GPU, and NPU, with Intel’s AI Boost silicon contributing 11 TOPS to the total.

That’s significantly better than the ARM chips, although compared with the likes of the Nvidia high-end GPUs, like the RX 5090, which can muster 3,352 TOPS, it's still at the modest end of the scale. However, this NAS is sufficient for running local AI models, and using the SFF-8654 port, external GPUs can be added to significantly bolster AI capability.

Overall, the hardware in the D6 Ultra is impressive, even if Intel has released better chips since the Meteor Lake era.

• Features: 3.5 / 5

Zettlab D6 Ultra: Software

• ZettOS
• ZettAI

Since the likes of Synology, Qnap and Asustor all have mature NAS operating systems, the best comparisons can be made between ZettOS and Ugreen’s evolving UGOS Pro operating system.

My immediate reaction to ZettOS was that even in this early stage, it has features that took at least six months or longer to appear on UGOS Pro, and a much better app selection.

These include support for Docker and Virtual Machines, media tools, Home Assistant, Plex, Jellyfin, Unifi, and a collection of developer tools.

While I’d have expected to see more software development options, the inclusion of Docket and VM provides an easy means to add those things either with a container or a VM of a desktop Linux distro.

I’m not a huge fan of the red, yellow, green dots for window controls, aping Apple; the Web interface is relatively clean and doesn’t require supporting documentation to navigate.

But there are a few significant holes in the feature selection of the OS, most notably with respect to security. At the time of writing, there is no 2FA, limiting access to the machine via a login and password, and if you use the Windows Zettlab AI NAS app, those are both stored on the client PC. It is possible to use a Zettlan Remote ID to connect to the NAS externally, using the Zettlab cloud portal.

I was a little shocked by some of the security choices made for this unit when I discovered that by default, the FTP server functionality was active. That’s not typically considered a wise move, and admins only activate that feature when they’ve put in place controls to avoid it being externally exploited.

Another area where this NAS OS veers slightly off the beaten path is its file system, which, to my understanding, is a proprietary one developed by Zettlab. Those expecting the choice of Ext4 or BTRFS will be disappointed, and I don’t think the current file system supports a hybrid structure with drives of different sizes. For sharing, SMB and NFS are supported, but I didn’t see any means to format USB-connected storage. In fact, all the external drives I connected, either to USB 3.2 or USB4 ports, were ignored. Eventually, I got a thumb drive that appeared to be formatted in FAT32, but drives that were preformatted in exFAT or NTFS were not recognised. That’s a feature that needs to be made a priority, I’d suggest.

Having two 10GbE LAN ports offers some great network bandwidth, but there currently aren’t any link aggregation or failover options to leverage the full potential of them.

In my tests, the USB4 ports did not work in host mode, although this feature, I believe, is promised. The USB-C ports did charge my laptop, at least.

The HDMI port does nothing currently, not even showing the Linux boot.

During my time with this machine, it underwent two firmware updates, suggesting that the software developers are backfilling functionality that’s either missing or not working optimally.

• Software: 3.5/5

Zettlab D6 Ultra: Performance

• Network performance
• External drives limitation
• AI models

For my testing, I used six IronWolf 4TB drives and allocated them into a RAID 10 pack for the best possible speed available. And to enhance that further, I allocated one Crucial P3 NVMe drive as a 1TB cache.

Like with many NAS, for whatever reason, a single drive can only be allocated to caching reads, and it takes two modules to cache reading and writing. I didn’t have two spare M.2 drives, so I went with the cached reading instead.

Over a single 10GbE, the read throughput hit over 900MB/s, which is excellent. As there is no link aggregation on the network ports, that’s realistically as fast as it’s possible to go. Should host mode on the USB4 be made active, that should be capable of much faster speeds. But without caching, six hard drives hit a bandwidth ceiling of around 900MB/s, since each is only capable of about 150MB/s. For this reason, unless you run SATA SSDs or have large M.2 cache drives for both reading and writing, there is little point in using the SFF-8654 port to add more or faster network ports.

And, in the support material, it states that “We currently do not natively support U.2 or U.3, but our machines can expand via the SFF-8654 interface.” Since SFF-8654’s function on some motherboards is to connect U.2 or U.3, that seems an odd choice.

As I mentioned earlier, external drives' functionality is incomplete, and without support for NTFS and exFAT, it's extremely limited in what it can be used for. I also found it disappointing that if the system didn’t recognise the file format, it didn’t offer to format it into one it was happy to work with.

That fun was as if nothing compared to the adventure of using this NAS as an AI local platform.

Inherently, the D6 Ultra and D8 Ultra support a local LLM model that can analyse whatever documents you put on the NAS, providing a chatbot interface to work with the contents.

If that sounds great to you and you have lots of files you need to navigate with AI logic, then right out of the box, this might be the NAS for you.

When the processing of the local model occurs is configurable, so that it doesn’t step on current tasks like file serving.

What’s great about this functionality is how automatic some of it is. After loading some prior review content folders onto the NAS, I discovered that the LLM had created a photo album based on the files and the subjects that it saw in the images. It could then also answer questions about the files, revealing the knowledge it has gained processing them.

Or rather, that’s what is implied. Except when you ask it in the AI chat window, it wants you to specifically say which files it should check, which isn’t super-helpful.

I should also say that the default ZettAI created by Zettlab, I assume, was poor at some general AI tasks, like history.

To further explore this feature, I looked at all the models that the system has available to install. These included four variants of Gemma, the Google AI, four flavours of Phi, the open-source Small Language Models (SLMs) created by Microsoft, two more QwQ models made by the Owen Team, two DeepSeek-R1 models, and a couple of Meta-made Llama models.

It’s possible to load and use each of these, though they range in size from about 2GB to more than 4GB, and some use plenty of memory.

I tried a number of them, and to put it mildly, my mileage varied considerably.

What I learned was to not ask DeepSeek about history, since it failed the most basic questions about historical events.

Its collection of the Kings and Queens of England was horrifically wrong, with it deciding that Queen Elizabeth I reigned from 1237 to 1558, which would have easily made her the oldest person ever recorded, if it were true. For those wondering, her reign lasted from 1558 to 1603.

Realizing what a rich mine of alternative information DeepSeek could be, I then asked about which US presidents died in office, and it completely messed up that challenge. It said ten presidents had died while in office, whereas the right answer is eight. It got the names of those eight wrong, included people who died after they left office, and insisted that three presidents died while hunting.

ZettaAI did a better job of the King's question, although not perfect, as it left out Harold, who died at Hastings. But it entirely messed up the dead Presidents, leaving out Zachary Taylor, Abraham Lincoln, James A. Garfield, Warren G. Harding and even John F. Kennedy. Then it included Andrew Jackson, Martin Van Buren, Theodore Roosevelt, Woodrow Wilson, Dwight D. Eisenhower, Lyndon B. Johnson, Richard Nixon and Gerald Ford, none of whom died in office. Out of eleven U.S. presidents it provided, the only ones it got right were William Henry Harrison and William McKinley.

I’m sure it’s possible to find subjects that these AIs are much better with than Western history, but the point is that if you were a student using these tools for homework exercises, you could be in deep, deep trouble.

To be clear, the effectiveness of these models, or not, isn’t a reflection on the Zettlab D6 Ultra, but the nature of AI technology, and its value to those expecting it to come back with generally correct answers.

When I questioned the ZettAI about how Abraham Lincoln wasn’t in the list of Presidents in one of those who were assassinated, it tried to say that the list of those who died in office didn’t include those who were killed, even though that was not a context I created. Then contradicting itself, it also argued that Lincoln died of Pneumonia, caused by the gunshot, but not directly from the assassination. This appears to be a riff on the concept that guns don’t kill people; complications from gunshot wounds kill people.

In short, if you are expecting something as powerful as datacentre AI in a small box on your desktop, you might need to scale that objective back that thinking somewhat, though as models improve, it might become an invaluable tool.

• Performance: 4 / 5

Zettlab D6 Ultra: Final verdict

There are some positive things to say about the D6 Ultra, since, for a NAS platform in the earlier stages of development, ZettOS is already reasonably sophisticated.

Where more questions exist is in the value of AI on hardware like this, because, as nice as a Core Ultra 5 CPU is, it's hardly a data centre. Depending on the size of the dataset you wish to use with the AI, this could be a highly responsive and productive solution, or something painfully slow to access.

I’m aware that even more powerful NASs are coming along that can outperform the D6 Ultra, although if an external GPU were added to this platform, it might be quicker. What’s not a guess is that if you added an RTX 5090 to this machine externally and gave it 96GB of RAM to run its models, this could be an impressive local AI solution, but the system's cost would be a minor part of the total expenditure.

Given the power of datacentre AI solutions, a solution like the D6 Ultra is likely to interest only those who want to use models experimentally or isolate the development of an AI platform from the Internet.

But I have to question how cost-effective this would be in the long term, should the model reach a level of complexity where the NAS struggles to run it interactively to achieve the level of performance you might want.

Should the AI bubble burst and people realise that it's of limited use for many tasks, at least this hardware is sufficient to be an excellent file server and media system.

Should you buy a Zettlab D6 Ultra?

Buy it if...

Don't buy it if...

For more NAS solutions, these are the best NAS hard drives we've reviewed.