When OpenClaw hype was at its peak in the beginning of 2026, a lot of people bought Mac Mini just to keep an agent running around the clock.

I never understood why people would spend $500 on a device to run OpenClaw like agents that are merely inferences to a commercial LLM like Claude and ChatGPT.

But these 'claws' are here to stay, and people are utilizing them to run their agent on dedicated hardware. And that's why we have dedicated hardware built around these agents.

Autonomous is a company known for standing desks and ergonomic chairs, although they did start as Robotbase, making personal robots.

Seems like they are back to their roots, sort of. They have a few "AI hardware" projects in their offering and that includes a thinking desk and a paper pet companion.

But the device I am going to feature today is a desk companion that runs AI agents and you can use it as an office assistant or an intern (no offense to interns).

The device is called Intern 2 because it's a follow-up to Intern, which was released six months ago. Let's learn more about it.

Intern 2 Specifications

• Board: Orange Pi 4 Pro, 8-core Allwinner A733 (2x Cortex-A76 + 6x Cortex-A55, up to 2.0GHz)
• Memory: 6GB LPDDR5
• Storage: 64GB onboard + up to 256GB expansion
• Connectivity: Wi-Fi 6, Bluetooth 5.4
• New in Intern 2: voice interaction (dual mic, pickup up to 3m; 5W speaker) and browser control, on top of the existing Telegram/Slack/Discord channels
• Warranty: 1-year warranty, 30-day free returns

Intern 2 starts at $249 for the OpenClaw or Hermes editions, and $299 for the Developer edition.

So basically, the device ships with an agent framework already running. You pick OpenClaw or Hermes at checkout, plug the box into Wi-Fi, add it to a Telegram, Slack, or Discord group (WhatsApp gets added later through a separate QR pairing step), and it starts triaging your inbox, watching GitHub issues, and running whatever multi-step task you hand it, all without you touching a terminal.

I am guessing you will have to do some configuration to connect it to your mail and suff but that should not be too difficult if your main LLM is already connected to various services you use.

The board underneath is an Orange Pi 4 Pro: an 8-core Allwinner A733 chip, two beefier Cortex-A76 cores paired with six efficiency Cortex-A55 cores, and a 3 TOPS NPU. It's the kind of board you can buy separately, so Intern 2 isn't offering or hiding anything exotic. It is just packaging familiar SBC hardware into something that doesn't look like a dev board.

What does it look like? It looks like a pyramid with a glowing top. I kind of like the look.

The voice interaction is the main selling point here. Why type when you can just speak to it? Is that not what Tony Stark did as Iron Man?

Intern 2 adds dual far-field mics rated to pick up speech from up to 3 meters away, plus a 5W speaker, so you can just talk to the thing sitting on your desk the way you'd talk to a coworker (or an intern). Browser control is also there, letting the agent act on websites on your behalf instead of only working through chat integrations and APIs.

There are three editions to choose from:

• OpenClaw ($249): the default, framework preloaded, no setup beyond Wi-Fi and a chat app.
• Hermes ($249): identical hardware, running the Hermes agent out of the box instead.
• Developer ($299): ships with the shell open. You SSH in, and Claude Code comes preloaded, with room to swap in Codex or OpenCode if you'd rather run your own agent stack.

Open-source Autonomous OS

Intern 2 runs on Autonomous OS, an open-source operating system created by Autonomous for physical AI agents.

It handles sensing, memory, skills, and updates, while the actual "brain" is swappable between OpenClaw, Hermes, PicoClaw, Claude Code, OpenAI Codex, or any other agent framework you want to wire in.

Intern is the second reference device built on it, after a desk lamp called Autonomous Lamp. Since Autonomous is working on more AI-powered hardware, it makes sense to have a custom operating system.

Open sourcing is also the wise decision. These devices are more for tech enthusiasts, and some of them may tweak the OS to our liking and needs. So good job, Autonomous.

Pricing and Availability

Intern 2 is available right away.

As mentioned previously, it starts at $249 for the OpenClaw or Hermes editions and $299 for the Developer edition, with free shipping. Every unit ships hand-assembled and tested, backed by a 1-year warranty and 30-day free returns (I like that). Developer documentation is available here.

Should you buy Intern 2 is a question only you can answer. If you have interacted with Nanoclaw, Hermes Agent, etc. and like the idea of interacting with them via voice, this device should interest you.

Sure, this can be built from scratch, especially when the operating system is also open source. But not everyone has the time and skills to do that. People spend money on 'out of the box' experiences after all. I would not mind splurging on it, as I am sucker for such gadgets.

So, would you spend some money and get yourself an "AI Intern"?

Ever tried a DAS (Direct-Attached Storage) device?

Unlike NAS, where you have a bunch of hard drives available over network, DAS is directly connected to your computer.

I never used such a product because I never had the need. But then Orico sent me there new DAS 88 Series 4-Bay USB4 NVMe SSD Enclosure and I got the opportunity to use a DAS for the first time ever.

Sharing my experience and some benchmarking test that I did for the first time in several years.

Orico 8848U4 DAS Specification

Here's the quick hardware specifications for Orico 8848U DAS:

Do note that there is also a SATA version of DAS in the same 88 Series lineup. It is older, bigger and has a RAID mode. The version I tested is M.2 NVMe SSD only and without built-in RAID functionality. That's intentional to give you the raw NVMe speed in a compact form factor.

Do note that this is a diskless enclosure. You bring your own M.2 NVMe drives. You cannot expect to get enclosures with SSD disks for under $200 in this age of AI slop.

Build Quality And Design

The first thing I noticed is how good this thing looks. The CNC-machined aluminum body has a silver finish that would blend right into a Mac-heavy studio desk. I am a Linux person, not an Apple person, but I will admit the build quality is quite impressive. Feels solid. My Terramaster NAS also has silver-gray aluminium chasis but this one is more "Mac looking."

All the ports and controls are on the back. There is a USB4 Type-C port, 12V DC power input, a dedicated power button, and a cooling fan.

The power button being on the back is a minor annoyance if the enclosure is tucked away. Worth planning your desk layout around it. The power button glows when pressed but difficult to look at it in the back.

The front has disk bay access with a slider button. It's a little bit stiff but not worth complaining about. The disk bay can host M.2 NVMe SSDs of various sizes. I only tested it with 2280 but guess that doesn't really matter.

The front also has 4 indicators at the bottom. They glow blue when the DAS is attached to a running computer.

Sides has nothing, just minimal branding.

The fan is there for active cooling, and ORICO claims it operates under 30dB. In my testing, the device felt completely silent. The fan does start running as soon as power is turned on but there is no audible noise unless you put your ear near the fan vent.

There is no fan speed control that I could find, which is a small omission. It would be nice to have a quiet mode toggle, but in practice it's a non-issue.

Keep in mind that the enclosure has no passthrough port. There is a single USB4 connection and that's about it. If you need to daisy-chain other Thunderbolt peripherals, this will consume your only port. On a laptop with limited Thunderbolt ports, this is something worth thinking about before buying.

Linux Compatibility

I noticed that the drive appeared as a native NVMe device. This means that each drive in the enclosure shows up as a separate NVMe namespace (nvme1p1, nvme1p2 etc) under a single controller.

$ cat /sys/class/nvme/nvme1/transport
pcie                             #output

And this detail matters because it means that the Thunderbolt PCIe tunneling is working correctly. The OS treats these drives as if they are plugged directly into the motherboard's PCIe bus, not as USB devices. My Sandisk external SSD comes up under /dev/sda that means it's treated as USB.

Performance Testing

I tested with a Crucial CT500P3PSSD8 (P3 Plus 500GB NVMe) installed in one bay. I only had one drive available for testing. Rest of them were being used in ZimaCube and Terramaster NAS. I have to buy new NVMe SSDs but the prices are not coming down. I ran benchmarks using fio and hdparm, and also timed some real-world file transfers.

Understanding The "40Gbps" Claim

It is easy to get excited with numbers but let's analyze the numbers. ORICO itself clarifies this in the product page: each drive achieves up to 10Gbps when all four bays are operating simultaneously. The 40Gbps is the total aggregate across all four drives. A single bay doesn't get 40Gbps.

In practice, even that 10Gbps per drive is rare achievement. Thunderbolt 4 tunnels PCIe 3.0 x4, which gives around 3500 MB/s of usable bandwidth total. Divided across four bays, each drive gets roughly 800-900 MB/s of real-world headroom. Mind the difference between bits and bytes.

My single-drive benchmarks came almost there, which indicates that the enclosure is performing exactly as it should. If I had 3 more spare SSDs, I would have tested the full 40Gbps claim.

Benchmark Results (ext4)

I tested with two filesystem configurations. The reliable numbers come from the ext4 run, where direct I/O (O_DIRECT) was fully confirmed working with no cache assistance.

Sequential read at 729 MB/s and sequential write at 669 MB/s are pretty good numbers. Pretty close to the numbers I discussed earlier. That's the advantage of using NVMe over a PCIe tunnel.

NTFS vs ext4

Since the Crucial drive I installed was previously formatted as NTFS, I benchmarked it in that state before reformatting to ext4.

NTFS sequential read was partially served from OS cache due to FUSE limitations and thus it is not a reliable number. Don't think that NTFS is somehow better than ext4 ;)

The NTFS write performance is inconsistent. Sequential writes drop to 231 MB/s and random 4K writes fall to around 2.3 MB/s.

Mind that this is not the DAS's fault. Linux accesses NTFS through the ntfs-3g FUSE driver, which adds overhead, especially on small random writes. The raw device speed (hdparm on the block device directly, bypassing any filesystem) was virtually identical in both runs at around 764 MB/s, which confirms the ORICO hardware is not the bottleneck.

Real-World Transfers

I am not a fan of bechmarking tests. They do not capture how the device actually feels to use. So I ran two practical tests.

Copying a 5GB 4K video file to the DAS took just over 4.5 seconds.

A folder with nearly 2,800 photos and videos totaling 10.2GB transferred in under 12 seconds.

Note that Linux buffers writes in RAM before flushing to disk in the background, so these numbers reflect the immediate user experience rather than sustained disk throughput.

Things To Keep In Mind

The no-passthrough situation is probably the biggest practical limitation but only if you need to daisy chain multiple devices and only have few thunderbolt ports available on your system.

No RAID support is by design here, but something to keep in mind if you were hoping for redundancy.

Fan control would be a nice addition, perhaps? The device was silent in my use, but there is no software or hardware toggle to set a fan curve or force it off. Definitely not a dealbreaker, just something I noticed.

Another thing to note is that there is not enough space for a heatsink since the device has a compact size.

Also note that for $219 you are buying a diskless enclosure. Four decent NVMe drives to fill it will cost a lot more. That's an obvious thing but I still would like to state that.

Who Is This For Really?

The primary userbase for something like this is video editors and creative professionals who are constantly moving large files.

If you are working with 4K or 8K footage across multiple projects, you'll have TBs of data and your internal SSD would fill up fast. A 4-bay NVMe DAS sitting on your desk gives you that additional storage without the latency of a NAS. You plug it in and it just works like local storage, because over Thunderbolt 4, it effectively is.

It is also a good fit for you if you already have a handful of spare NVMe drives sitting around from previous hardware upgrades. Instead of those drives collecting dust in a drawer, you slot them in here and have a fast, compact multi-terabyte storage pool at your desk.

That said, this is not for everyone. If you just need an extra 1 or 2TB of portable storage, a simple USB-C external SSD will do the job at a fraction of the cost and with far less desk space and will also be portable. A dedicated 4-bay DAS only makes sense when you need the multi-bay capacity.

Some NAS devices offer a feature called direct attach, where you plug the NAS directly into your computer over Thunderbolt and use it as local storage rather than over the network. ZimaCube highlights this as a use case.

The difference here is that a NAS with decent specs will cost you significantly more than this DAS. If all you need is fast desk-side storage without the NAS software stack and network overhead, the ORICO DAS is a simpler and cheaper solution.

Conclusion

As a Linux user, the experience was better than I expected. No driver issues, no special configuration needed.

The drive showed up as native NVMe devices and behaved like internal storage. Format to ext4 and you get fast sustained speeds in both directions.

Do evaluate your needs and if it fits, you can get it from its official website or order it on Amazon:

I already reviewed the original ZimaCube Pro and gave it a positive verdict, for the most part.

So when IceWhale offered me the newer version for review, I was curious to see how much had actually changed.

Not much specification-wise. Here's why. If you compare the spec sheets of the ZimaCube Pro (original) and the ZimaCube 2 Pro side by side, you will find the same Intel Core i5-1235U processor, the same 16 GB DDR5 RAM, the same 256 GB system SSD, and the same six SATA bays. The core hardware of the Pro tier has not changed between generations.

What IceWhale actually did with the ZimaCube 2 lineup is more subtle. The real generational leap happened at the entry level of ZimaCube. The new $799 Standard model replaces the old N100-based unit with a proper Intel Core i3, Thunderbolt 4, and DDR5, which is a substantial and much needed step up. Intel N100 was not the best choice even two years ago.

The Pro, meanwhile, got refinements, especially on the cooling and fan noise side, along with a more mature ZimaOS. The Pro version is not a new machine, it's just a better-tuned version of the same one.

With that said, here is everything I learned from almost a month of use.

ZimaCube Pro v2 Specifications

The ZimaCube 2 is available in three models:

• ZimaCube 2 Standard ($799): Intel Core i3-1215U, 8 GB DDR5, 256 GB system SSD, 2.5GbE networking
• ZimaCube 2 Pro ($1,299): Intel Core i5 (12th Gen), 16 GB DDR5-4800, 256 GB PCIe Gen4 SSD, 10GbE + 2.5GbE networking
• ZimaCube 2 Creator Pack ($2,499): Everything in Pro, plus NVIDIA RTX Pro 2000 GPU, 64 GB RAM, 1 TB system SSD

The unit I have for this review is the Pro, and I added an NVIDIA RTX 2000 Ada (16 GB VRAM) via one of the PCIe slots bringing it close to Creator Pack territory. This gives local AI capabilities to my otherwise Homelab device.

Again, the Pro specs are essentially identical to the original ZimaCube Pro. The i5-1235U, 16 GB DDR5, 256 GB SSD all unchanged. The ZimaCube 2 lineup's real generational upgrade happened at the base model ($799 Standard with i3), not the Pro tier.

Key specs for the Pro version:

• CPU: Intel Core i5-1235U, 10 cores
• RAM: 16 GB DDR5-4800 (expandable to 64 GB)
• System Storage: 256 GB PCIe Gen4 x4 SSD
• Drive Bays: 6x SATA (3.5"/2.5") + 4x M.2 NVMe slots
• Networking: 10GbE + 2.5GbE LAN (no WiFi)
• Ports: 2x Thunderbolt 4 (USB-C, 40Gbps), 2x USB-A (front), 2x USB-C 5Gbps + 2x USB-A (back), DP, HDMI and audio jack
• Expansion: 2x PCIe slots (4.0 X4 and 3.0 X2)
• OS: ZimaOS (default), but you can install Proxmox, Unraid, TrueNAS, Ubuntu, and others

What else in the Box?

The accessory bundle is generous and thoughtful. You get:

• Four heatsinks for the M.2 NVMe drive slots
• A couple of screwdrivers so you can actually open and configure the device without hunting for tools
• Plenty of screws for hard disks and SSDs
• A high-quality Cat6 UTP cable for networking
• A power cable matching your region (a small but sensible touch since they ship worldwide)

Fan noise is handled a lot better now

One of the major complaints I had with the previous ZimaCube was the fact that the tower cooler fan was annoyingly loud. I could not work without my Bose QuietComfort headphones on. I am a little picky about keeping the work environment quiet.

ZimaCube Pro v2 is noticeably quieter than the original generation. That is a real improvement and definitely worth acknowledging.

It is still not an absolute silent device. There is still a faint, constant, low-level whirring from the fans that you may hear in a quiet room if you are standing too close to it and the device is under load. You'll only notice it if you want to notice it. My TerraMaster NAS operates in near-silence by comparison.

Where you place this device matters. A dedicated server spot away from your workspace: no problem. On your desk in a home office, you might notice it.

The ZimaOS Experience: Still the heart of this device

ZimaOS is the backbone of ZimaCube and all other Zima devices. The good news is that it has matured considerably since the original ZimaOS that was released with the previous ZimaCube.

The interface is familiar: a clean web UI accessible from any browser on the local network, a file manager, and an app store for one-click deployment of self-hosted software. The applications run in Docker containers underneath, but you do not need to know Docker to install or use them.

The storage manager continues to be one of the strongest features.

Disk management, merging storage pools, and setting up RAID are all handled through a graphical interface that makes things a lot easier.

I migrated two drives from my original ZimaCube: a 1 TB SSD and a 500 GB SSD.

A note on RAID for anyone in a similar situation: yes, ZimaCube 2 supports RAID and the ZimaOS GUI makes creating a RAID array incredibly easy.

But with mismatched drive sizes, RAID is wasteful. A 1 TB + 500 GB RAID 1 mirror gives you only 500 GB of usable space, wasting the rest.

So I skipped RAID and went with a practical split setup instead:

• 500 GB SSD dedicated to app data (persistent storage for all Docker-based apps)
• 1 TB SSD for actual user data and files

Moving the application and container data between disks is effortless thanks to the storage manager.

ZimaCube has 256 GB of built-in storage. But it starts filling up quickly as you install more (Docker) applications and add user data.

This is why I moved the app data to a 500 GB SSD and the user data to a 1 TB SSD. This way, the built-in storage is free. Once migrated, applications like Immich and others use the dedicated drive without needing to manually configure storage volumes in every app.

ZVM (Zima Virtual Machines) is also available, letting you run other operating systems directly on ZimaCube from the browser. It comes with a trial of Windows in a VM by default. Not useful to me but I guess it extends the functionality and offering of ZimaCube.

ZimaOS has many little gems that are not well advertised. For example, there is an option to enable UPS settings. In the event of a power failure, if you have added a UPS in the set up, you can choose to gracefully power off the ZimaCube if the power is not restored in the given time frame. This saves the disks and data from a prolonged power failure.

Many self-hosted applications are required to be configured with username and password at installation/deployment time. In such cases, ZimaOS shows the default credentials used under the Tips section. It is accessible when you click on the three vertical dots on the app.

ZimaOS still has room to grow!

The GUI hits its limits more often than the marketing suggests. Some things still require dropping into a terminal, which is fine for Linux users but undercuts the "easy homelab for everyone" pitch for less technical users.

For example, I installed Ollama-Nvidia from the app store. But there was no further information on using it. Since I have used Ollama in the past and I also have experience with ZimaOS, I figured out that it has to be accessed via the terminal of the Docker application.

When I ran Ollama commands, there was no visible output while the model downloaded or installed. It appeared to do nothing and then it was done. For anyone unfamiliar with CLI tools, this will cause confusion.

Thankfully, Open WebUI now directly downloads models from Ollama so a separate Ollama server is not mandatory anymore. But for agentic AI flow, I think Ollama is still better suited.

Jellyfin as a media server

While ZimaOS does makes self-hosting open source software easier, individual applications do require occasional tweaking and configuration changes.

For example, I installed Jellfin Nvidia GPU version as I have added the RTX 2000 Ada. When I tried streaming a 4K video file, the GPU was not used at all. It started consuming 90% of the CPU.

It's because Jellyfin still needed configuration changes and I had to enable the hardware transcoding with NVIDIA NVENC.

And once I did that, the CPu consumption was at the minimum. GPU was consumed but not much as NVIDIA RTX 2000 Ada is way too powerful for video transcoding.

The point I am trying to make here is that individual apps may require setting and configuration changes of their own.

Zima has improved its documentation considerably and mostly covers such pain points but there is plenty of scope for improvements here.

Photo storage and management with Immich

Immich is a self-hosted Google Photos alternative and ZimaCube is an excellent host for it. The combination covers the full use case: mobile photo backup via the Immich app, face recognition, smart albums, CLIP-based semantic search, and a clean timeline interface in the browser.

With a GPU present, Immich's machine learning tasks like facial recognition, smart search run noticeably faster than on a CPU-only setup. I forgot to benchmark this data. Sorry for that 😦

The ZimaCube mobile client provides basic folder sync as an alternative if Immich feels like too much setup for your needs. Simple photo backup from your phone can be done without any additional app configuration. Although, it was buggy in my experience and I would prefer relying on PhotoPrism instead.

Using ZimaCube 2 for local AI

This is where ZimaCube Pro v2 genuinely separates itself from a typical NAS as no traditional NAS manufacturer gives you a PCIe slot for a GPU. ZimaCube does, and that changes what the machine can be.

I shared my experience with Ollama earlier but then you don't necessarily need Ollama all the time. Open WebUI, AnythingLLM works well too.

Anyways, the runtime experience is a different story. Once Ollama is running and models are pulled, things are smooth. With the RTX 2000 Ada's 16 GB VRAM, 7B and 13B parameter models run entirely in VRAM with no memory spillover. Response times are fast enough for real use: comfortable for chat, fast for summarization tasks, very fast for coding assistance.

For everyday tasks like text summarization, Q&A, quick coding help, 7B to 13B models are the practical sweet spot. They are fast, fit cleanly in 16 GB VRAM, and the quality is more than adequate. Larger models in the 32B+ range will start to push the memory limits, though quantized versions can still run well.

I tried Deepseek-r1, qwen and llama primarily. I wanted to use Kimi but Ollama only has it in cloud mode and that fails the point of setting up a local AI.

I used AnythingLLM as an AI assistant for my private notes. It helps me find and create quick tips for the newsletter and social media. Setting up AnythingLLM was a struggle too.

On the agentic AI side, I rely on Nanoclaw as it is super easy to setup. ZimaOS only offers OpenClaw which is more popular but not something I prefer. Setting up Hermes Agent is easier as it has graphical interface.

Let's be clear. Without a GPU, local AI on ZimaCube is CPU inference. Functional but slow and limited to smaller models. The PCIe slot is what makes the real difference here but then you have to spend on a small form factor GPU as not all GPUs will fit in the ZimaCube.

Linux is still not getting the place it deserves

There were two main concerns that I raised in the previous ZimaCube review: it's fan was too noisy and the Zima client was not available for Linux.

Built on ZeroTier, Zima client allows you to use your ZimaOS device from anywhere in the world. You have to enable remote access option for this.

This way, even if your ZimaCube is on your home network, behind a NAT, you can access it from another location, city or country. This makes your homelab available outside the home network.

The problem is that you need to install the Zima Client application on your system and this application is still available only for Windows and macOS.

Two years and we still don't have Zima Client for Linux and as a Linux user, I am surely disappointed.

What I am using ZimaCube Pro v2 for

I have a few Zima devices in my setup. I started my homelab journey with the ZimaBoard, a smaller SBC type of device that came with CasaOS. It was upgraded to ZimaBoard 2 and it runs continuously in my setup and powers my Home Assistant and Jellyfin servers.

ZimaCube 2 runs the same operating system and thus I could use it for all my homelab need and remove ZimaBoard from the picture, but I chose not to do so.

I use ZimaCube 2 as a combination of NAS and local AI. It has more power and thus also consumes more power. I usually turn the ZimaCube off at the end of the day and keep it running during my working hour.

Photo backup and management with Immich is the primary use case. My full photo library lives here now.

Then I use Local AI inference with Ollama and Open WebUI. Running Qwen and a few other models for summarization, knowledgebase assistant helps (with Anything LLM on my laptop).

File storage for important documents. ZimaCube is one node in my backup strategy.

Is it worth getting?

ZimaCube 2 Pro is priced at $1299. There is an increase in price because SSDs, RAMs cost more (all thanks to the AI craze).

If you already own a ZimaCube Pro, there is genuinely no hardware reason to upgrade. The processor, RAM, and storage are the same. It is not worth the cost of switching just for a more silent fan.

If you are buying for the first time, the ZimaCube 2 Pro is the one to go for. The standard ZimaCube 2, priced at $799, also provides more value now.

Some people may think that it is better to build a PC in the same or cheaper price but than not everyone would want to do that. Devices like ZimaCube serve a niche of people who want to own a homelab without spending too much time dealing with the hardware, networking and the operating system.

These days, you need more than just a NAS, You need that hardware to do more than just storing data. And ZimaCube brings NAS, self-hosting and local AI together. It gives you PCIe expansion, Thunderbolt 4, and an approachable homelab OS in the same box. For the right kind of user, that combination has no real alternative.

For me, ZimaCube is an integral part of my home setup, and it stays that way.

P.S. You can win a ZimaBoard 2 (not ZimaCube 2) and start a small lab with that. Please enter the giveaway on X (formerly Twitter).