Minisforum's new flagship NAS comes with OpenClaw pre-installed — Strix Halo-powered N5 Max can run a local AI LLM

Minisforum N5 AI NAS — (Image credit: Minisforum)

Minisforum has announced an upcoming NAS that is built from the ground up to run large language models locally. The yet-to-be-released N5 Max AI NAS comes with a Ryzen AI Max+ 395 Strix Halo APU and features OpenClaw pre-installed, an open-source AI framework that can be configured to run a variety of tasks. Pricing and a release date have yet to be announced.

The small-form-factor manufacturer neglected to share the NAS's full specifications, particularly the unit's storage capacity. All we know officially is the CPU inside, which is AMD's flagship Strix Halo APU sporting 16 Zen 5 CPU cores that can clock up to 5.1GHz, a Radeon 8060S iGPU with 40 CUs, XDNA 2 NPU, and 64MB of L3 cache. The 395+ can be configured with 32GB to 128GB of system memory; likely, Minisforum is using a higher memory capacity of 64GB to 128GB. LLMs are known to scale very well with larger amounts of memory capacity.

Minisforum N5 AI Max — (Image credit: Minisforum)

That said, we can make some logical guesses about the system's other specs. Minisforum's Max variant of the N5 series appears to be using the same chassis as the outgoing N5 AI NAS and N5 AI Pro NAS. If this is true, the Max version will likely share the same storage configuration as the N5 AI/N5 AI Pro, consisting of five 3.5/2.5' HDD drive bays and three M.2 slots, two of which support U.2 drives. The HDD bays alone support up to 30TB per drive.

Article continues below

AI acceleration in network-attached storage systems is a trend that is growing rapidly in the computing industry. Having this capability in a NAS gives it the ability to serve as both a NAS and a local AI server for users. Having the LLM localized also improves security as all of the data processing and interactions are done within the machine and aren't shared with the internet.

OpenClaw is not an LLM like Copilot or Gemini, but is an AI framework that can be programmed to run a variety of tasks. For instance, OpenClaw can be programmed to run a photo search engine that can be controlled with conversational prompts. It can also be configured to edit videos based on prompts, automate emails, publish social media posts, and more. Specifically, OpenClaw routes messages to an LLM, which will then decide which tools to use to fulfill the user's request.

OpenClaw has exploded in popularity; however, security is one of the framework's biggest flaws. Beyond the apps' already problematic security issues that can leak sensitive data to the internet if not configured properly, malicious content has been found on ClawHub, a hub for OpenClaw users to install third-party extensions.

Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.

Aaron Klotz is a contributing writer for Tom’s Hardware, covering news related to computer hardware such as CPUs, and graphics cards.

6 Comments Comment from the forums

PEnns

"can run a local AI LLM"
"All we know officially is the CPU inside, which is AMD's flagship Strix Halo APU sporting 16 Zen 5 CPU cores that can clock up to 5.1GHz, a Radeon 8060S iGPU with 40 CUs, XDNA 2 NPU, and 64MB of L3 cache. The 395+ can be configured with 32GB to 128GB of system memory;"
Why does it need AI??

Are we talking about a NAS or a workstation, or what exactly???
Reply
CelicaGT

PEnns said:
"can run a local AI LLM"
"All we know officially is the CPU inside, which is AMD's flagship Strix Halo APU sporting 16 Zen 5 CPU cores that can clock up to 5.1GHz, a Radeon 8060S iGPU with 40 CUs, XDNA 2 NPU, and 64MB of L3 cache. The 395+ can be configured with 32GB to 128GB of system memory;"
Why does it need AI??

Are we talking about a NAS or a workstation, or what exactly???
So, I've been around the sun a fair few times and this may date me. Anyways when Compact Disc finally became popular every single audio product seemed overnight to gain some kind of stupid DIGITAL branding on it. DIGITAL everything, even on products where it was completely irrelevant and misapplied. I think perhaps, this is like that.. AI everything, even when it's not really. I mean yes/maybe in this case it but it can also be....a NAS, or a workstation? LOCAL AGENTIC AI!! Maybe just wipe it and install SteamOS. Yes, that one.

(Based on what I've read Agentic storage solutions could just maybe erase everything, but maybe it won't. It's a chance their willing to take with your data I guess)
Reply
alan.campbell99

Uh, nope. I just want a bunch of disks to store my stuff when I look at a NAS.
Reply
timsSOFTWARE

PEnns said:
"can run a local AI LLM"
"All we know officially is the CPU inside, which is AMD's flagship Strix Halo APU sporting 16 Zen 5 CPU cores that can clock up to 5.1GHz, a Radeon 8060S iGPU with 40 CUs, XDNA 2 NPU, and 64MB of L3 cache. The 395+ can be configured with 32GB to 128GB of system memory;"
Why does it need AI??

Are we talking about a NAS or a workstation, or what exactly???
It doesn't need AI, but one of the benefits of an APU like the Strix Halo AMD chip that they are using in this machine - or the Apple silicon chips - is that, as APUs, their graphics processor can use system RAM.

The problem that current regular PCs have for local AI hosting is that their architecture was built for traditional gaming and other graphics workloads, that require relatively little VRAM compared to system memory. But for AI inference - where you really would like to have the whole model loaded into VRAM, and decently powerful LLMs range in size from tens of GB to terabytes - that leaves you with the conundrum of having to choose between CPU-only inference (comparatively slow), GPU inference (fast, but you can only run small models), or rely on things like RAM or even disk offloading with MoE models (where weights are transferred back and forth between RAM and VRAM depending on which "experts" will be active for the next token - also relatively slow.)

The graphics processor on APUs with shared memory like AMD's Strix Halo and Apple silicon can access the entirety of the shared memory. So while they are typically slower than a traditional graphics card/have less memory bandwidth, it's a relatively affordable way to have accelerated inference with models that would be too large to fit on a consumer graphics card.

For example, you could run a ~70GB model on this machine, fitting comfortably in the 128GB shared memory, and get the benefit of APU-accelerated inference, whereas if you wanted to run something of the same size in GPU memory, you'd need at least a 6000 Pro Blackwell or several other cards (even though the 6000 Pro would be quite a bit faster).
Reply
USAFRet

Strike one brand off my short list to replace my 8 year old QNAP.
Reply
PEnns

timsSOFTWARE said:
It doesn't need AI, but one of the benefits of an APU like the Strix Halo AMD chip that they are using in this machine - or the Apple silicon chips - is that, as APUs, their graphics processor can use system RAM.

The problem that current regular PCs have for local AI hosting is that their architecture was built for traditional gaming and other graphics workloads, that require relatively little VRAM compared to system memory. But for AI inference - where you really would like to have the whole model loaded into VRAM, and decently powerful LLMs range in size from tens of GB to terabytes - that leaves you with the conundrum of having to choose between CPU-only inference (comparatively slow), GPU inference (fast, but you can only run small models), or rely on things like RAM or even disk offloading with MoE models (where weights are transferred back and forth between RAM and VRAM depending on which "experts" will be active for the next token - also relatively slow.)

The graphics processor on APUs with shared memory like AMD's Strix Halo and Apple silicon can access the entirety of the shared memory. So while they are typically slower than a traditional graphics card/have less memory bandwidth, it's a relatively affordable way to have accelerated inference with models that would be too large to fit on a consumer graphics card.

For example, you could run a ~70GB model on this machine, fitting comfortably in the 128GB shared memory, and get the benefit of APU-accelerated inference, whereas if you wanted to run something of the same size in GPU memory, you'd need at least a 6000 Pro Blackwell or several other cards (even though the 6000 Pro would be quite a bit faster).

I appreciate the explanation. Thanks!
Reply

Show more comments