Best Mini PC for Proxmox (2026): Ports, NVMe, RAM for VMs

Most “best mini PC for Proxmox” lists sort machines by price: budget, mid-range, high-end. That answers a question you probably have not asked. You are not choosing a price tier. You are choosing whether a machine can run the thing you intend to build.

Three things decide that. The Proxmox documentation sets the bar for two of them, and the way ZFS mirrors and boots sets the third: how many network ports the machine has, how many NVMe slots, and how much memory is left for your VMs once Proxmox, ZFS or Ceph has taken its share.

Quick picks

• Single node, a handful of containers: Beelink EQ14. Four N150 cores and a 16GB ceiling leave about 12GB for guests after Proxmox and the default ZFS cache. Two 2.5GbE ports mean it can still join a cluster later.
• Single node doing actual work: Beelink SER8. Eight cores and the lowest measured idle draw here. One network port, which is the catch.
• You want a ZFS mirror, or a cluster: Minisforum MS-01. Three NVMe slots, four network ports, a PCIe slot, and two 10GbE ports if you ever try Ceph.
• Heavy VM density: Minisforum MS-A2. Sixteen cores and thirty-two threads, with the running cost that implies.

If you want one answer: the MS-01 passes the port and slot rules below without also charging you for sixteen cores you may not need.

Check the MS-01 on Amazon↗

What Proxmox actually asks of the hardware

A few lines from the Proxmox VE documentation do most of the work in this comparison.

On networking: “We recommend a dedicated physical NIC for the cluster traffic.” The reason given is not bandwidth. Corosync, the layer that keeps cluster members agreeing with each other, is sensitive to latency jitter, and a link also carrying storage replication or VM traffic is exactly what can introduce it. The documentation adds that “a dedicated 1 Gbit NIC is enough in most situations” — so this is a question of how many ports you have, not how fast they are.

On cluster size: “If you are interested in High Availability, you need to have at least three nodes for reliable quorum.”

On the two-node case: “For smaller 2-node clusters, the QDevice can be used to provide a 3rd vote.” A QDevice is a small external daemon — a Raspberry Pi or an existing always-on box will do — that breaks ties. A second mini PC is the most natural upgrade path, and two nodes on their own cannot hold quorum when one goes down.

That turns two spec-sheet numbers into pass/fail rules:

• A machine with one network port cannot give corosync a dedicated link. It will still cluster, just in a configuration the documentation advises against.
• A machine with two NVMe slots cannot hold a ZFS mirror plus a separate boot device. You either boot from the mirrored pair or give up the redundancy you installed ZFS to get.

A mirrored pair for VM storage uses two slots. Proxmox can boot from that mirror, but then the disks holding your guests are also the disks holding the hypervisor, and replacing one touches both roles at once. Three slots keep those jobs apart: a small boot drive, and two disks that do nothing but mirror guest storage. Two slots force a choice, and the machine makes it permanently.

On the MS-01, the three slots are not equal. Minisforum lists them as PCIe 4.0 x4, PCIe 3.0 x4 and PCIe 3.0 x2, so the two drives in your mirror will not sit on identical links. Put the boot drive in the 3.0 x2 slot and the mirror gets the two x4 slots. On the MS-A2, the U.2/M.2 slot runs at PCIe 4.0 x4, and the other two run at PCIe 3.0 x4 by default and can be switched to 4.0 x4, according to Minisforum. So the MS-A2 can put its mirror on a matched pair.

How much memory is left for guests

Memory is the third rule, and it is the one the spec sheets overstate most. The Proxmox admin guide asks for “Minimum 2 GB for the OS and Proxmox VE services, plus designated memory for guests,” and adds that ZFS and Ceph need more on top.

ZFS takes its share without asking. On installs from Proxmox VE 8.1 onward, the ZFS documentation says “ZFS uses 10 % of the host memory, clamped to a maximum of 16 GiB” for its read cache, the ARC. Installs from before 8.1 used a larger default, so check yours. The installer writes that limit only when Proxmox itself goes on ZFS. If you boot from ext4 and add the mirror later, the ZFS default (62.5% on ZFS 2.3 and later) applies until you set zfs_arc_max yourself. For a single node with ZFS, the arithmetic is:

memory for guests = installed RAM − 2GB for Proxmox − 10% of installed RAM for the ARC

None of these machines reaches the 16 GiB clamp, which starts at 160GB installed. Memory sold as 16GB is 16 GiB, so the units line up. Worked through for each machine:

• EQ14 at 16GB (its one slot, full): 16 − 2 − 1.6 = 12.4GB for guests.
• SER8 at 32GB (the two 16GB modules in ServeTheHome’s review unit): 32 − 2 − 3.2 = 26.8GB. Beelink’s 256GB figure is not something two SO-DIMMs reach, so we do not work it through.
• MS-01 at 96GB (Minisforum’s figure): 96 − 2 − 9.6 = 84.4GB. At the 64GB Intel lists for the i5-12600H: 64 − 2 − 6.4 = 55.6GB.
• MS-A2 at 96GB: 96 − 2 − 9.6 = 84.4GB.

The 10% default is a starting point, not a recommendation. The ZFS documentation’s rule of thumb for the ARC is “at least 2 GiB Base + 1 GiB/TiB-Storage.” A mirror of two 4TB drives holds about 3.6 TiB, so the rule asks for about 5.6 GiB. On the EQ14 that leaves 16 − 2 − 5.6 = 8.4GB for guests. A Linux container can live in 512MB; a Windows VM or anything running a database wants 4–8GB to itself. So 8.4GB is one serious VM and a few containers. On the 64GB and 96GB machines, the default ARC (6.4GB and 9.6GB) already covers the rule for that mirror. The SER8 at 32GB is the exception: its default 3.2GB falls short, and sizing the ARC to the rule leaves 32 − 2 − 5.6 = 24.4GB.

Ceph on mini PCs: what the Proxmox guide asks for

Ceph is where these machines fall furthest from the documentation. The Proxmox Ceph guide asks for:

• “at least three (preferably) identical servers”
• “at least three nodes and at least 12 OSDs, evenly distributed” — four OSDs (drives) per node
• memory per OSD: “The OSD daemon requires 4 GiB by default,” and the guide recommends “at least 8 GiB of memory” per OSD for good performance, with more consumed during recovery
• a network of “at least 10 Gbps … exclusively for Ceph traffic”

Set against the four machines:

• Drives. None reaches four OSDs per node. The MS-01 and MS-A2 have three NVMe slots; with one booting Proxmox, two are left for OSDs, or six across three nodes. The EQ14 and SER8 have two NVMe slots, which leaves one OSD per node, or three in total. Even if every slot went to Ceph, three MS-01s would reach nine OSDs, not 12.
• Network. Only the MS-01 and MS-A2 have 10Gbps ports — two SFP+ each, so Ceph can have one to itself. The EQ14 tops out at 2.5GbE. The SER8 has one port in total.
• Memory. At 4 to 8 GiB per OSD, the EQ14’s one OSD takes 4 to 8GB of the 14GB left after Proxmox, leaving 6 to 10GB for guests. An MS-01 at 64GB with two OSDs gives Ceph 8 to 16GB and keeps 46 to 54GB for guests. At 96GB, the MS-01 and MS-A2 keep 78 to 86GB.

A three-node Ceph cluster of mini PCs is below the guide’s layout however you build it. The MS-01 and MS-A2 come closest: they meet the network rule and keep most of their memory. The EQ14 meets neither the drive count nor the network rule, and Ceph takes 4 to 8GB of its 14 spare gigabytes.

How we compared

Specifications come from the manufacturers’ own pages, and the memory and Ceph arithmetic from the Proxmox documentation. Two of the four published memory ceilings need a caveat, and the table footnotes give it. Beelink’s 256GB for the SER8 is more than its two SO-DIMM slots can hold. Minisforum’s 96GB for the MS-01 is higher than the 64GB Intel lists for the i5-12600H; the table shows Minisforum’s 96GB and footnotes Intel’s 64GB.

Idle power comes from published independent measurements, each with the conditions it was taken under. Those conditions differ: the EQ14 figure was taken under Windows and the MS-01 figure under Ubuntu, and the table notes say which is which. Three-year electricity uses the idle figure at the U.S. average residential rate published by the EIA. No dollar prices are listed — the table shows a rough price band, and the buttons go to the current Amazon listing.

We do not own these machines. Everything below is drawn from published specifications and from reviews by people who measured them.

The machines

Beelink EQ14 — the cheapest way to get two ports

The N150 is a four-core, four-thread part in the 6W class, and the EQ14 pairs it with a single DDR4 SO-DIMM slot capped at 16GB. That ceiling is the story: about 12GB for guests with the default ZFS cache, and about 8GB if you size the cache for a mirror.

What is easy to miss is that it has two 2.5GbE ports. At this price, that is the difference between a node that can eventually join a cluster properly and one that cannot. If the plan is “one small always-on box now, maybe a second later,” this is the cheapest way to keep that option.

The one published idle measurement is igor’sLAB’s 11.8W, taken under Windows rather than Linux, and Beelink does not publish a figure. That is above the SER8’s 7–10W, so the EQ14’s case is the purchase price and the second port, not the power bill.

Check price on Amazon↗

Beelink SER8 — the efficient one with one port

Eight Zen 4 cores, sixteen threads, DDR5, and the lowest measured idle draw in this group: ServeTheHome measured 7–10W at the wall. Three years of electricity at that draw comes to $34 to $48.

Two things need saying plainly.

It has one 2.5GbE port. For a single node, one port is all you need. But the SER8 cannot follow the documentation’s networking recommendation if you later cluster it, and the upgrade path is a USB adapter — exactly the kind of link you do not want corosync depending on.

Its memory ceiling is not what the number suggests. Beelink publishes a maximum of 256GB. The board ServeTheHome reviewed has two SO-DIMM slots, fitted with two 16GB modules, and two SO-DIMMs do not add up to 256GB at any capacity sold today. What you can install is the largest pair of SO-DIMMs you can buy.

Check price on Amazon↗

Minisforum MS-01 — the one built for this

Three M.2 NVMe slots, two 10GbE SFP+ ports, two 2.5GbE ports, and a PCIe 4.0 slot. It passes the port and slot rules without argument: a boot drive plus a mirrored pair, and four network ports to separate corosync from everything else.

It is sold barebone — chassis, board and CPU — so memory and storage are on you. The memory ceiling has two published answers. Minisforum’s store page lists up to 96GB across two DDR5-5200 SO-DIMMs. Intel lists 64GB as the Max Memory Size for the i5-12600H. The table shows Minisforum’s figure and footnotes Intel’s; the memory section above works through both.

Power depends on the configuration. Scott Stuff measured 13.2W idle on an i5-12600H unit running Ubuntu 24.04 with no add-in network card. ServeTheHome measured 25–29W on their i9-13900H review unit. Which one describes your machine depends on the CPU and what you plug into that PCIe slot.

Check price on Amazon↗

Minisforum MS-A2 — density, and the bill for it

Sixteen Zen 5 cores and thirty-two threads, up to 96GB of DDR5, three NVMe slots, the same four-port networking as the MS-01, and a PCIe 4.0 x8 slot that supports splitting. Against the MS-01’s i5-12600H, that is four more cores and twice the threads.

ServeTheHome measured 25W idle, with one of their configurations idling at 45W. That is the widest spread in this comparison: at the U.S. average rate, the gap between 25W and 45W is around $32 a year, every year.

If you are consolidating several VMs onto one box, that may be the right trade. If you are running a handful of containers, you are paying for cores that sit idle.

Check price on Amazon↗

Compared on what Proxmox needs

Specifications from manufacturers. Idle figures from published independent measurements; see the notes for the conditions each was taken under. Power cost assumes continuous operation at the U.S. average residential rate, 18.31¢/kWh (EIA, July 2026).

Model	CPU	Max RAM (GB)	NVMe slots	Networking	PCIe slot	Idle draw (W)	3-yr power cost
Beelink EQ14Beelink	Intel N150 (4C/4T)	16	2	2x 2.5GbE	None	11.81	$57
Beelink SER8 (Ryzen 7 8845HS)Beelink	AMD Ryzen 7 8845HS (8C/16T)	2563	2	1x 2.5GbE	None	7-102	$34–$48
Minisforum MS-01 (i5-12600H, barebone)Minisforum	Intel Core i5-12600H (12C/16T)	965	3	2x 10GbE SFP+ / 2x 2.5GbE	PCIe 4.0 x16 slot (x8 electrical)	13.24	$64
Minisforum MS-A2 (Ryzen 9 9955HX, barebone)Minisforum	AMD Ryzen 9 9955HX (16C/32T)	96	3	2x 10GbE SFP+ / 2x 2.5GbE	PCIe 4.0 x8 (x16 slot, splitting supported)	25-456	$120–$217

Read the NVMe and networking columns first. They are the two that rule machines out. A second NVMe slot does not become a third, and one network port does not become two. Read the memory column through the arithmetic above: the number on the sheet is not the number your VMs get.

Which one, for what you’re building

Rows are what you are building, not what you are spending.

Use case	If the power bill matters most	If capability matters most
Single node, a few containers and VMs	Beelink SER8 (Ryzen 7 8845HS)	Beelink SER8 (Ryzen 7 8845HS)
Single node with a ZFS mirror	Minisforum MS-01 (i5-12600H, barebone)	Minisforum MS-01 (i5-12600H, barebone)
Two-node cluster (plus a QDevice)	Beelink EQ14	Minisforum MS-01 (i5-12600H, barebone)
Three-node cluster or Ceph	Minisforum MS-01 (i5-12600H, barebone)	Minisforum MS-A2 (Ryzen 9 9955HX, barebone)
Passthrough: GPU, HBA or a NIC	Minisforum MS-01 (i5-12600H, barebone)	Minisforum MS-A2 (Ryzen 9 9955HX, barebone)

The SER8 fills the top row and no other, and that placement is deliberate. Eight cores at the lowest measured idle draw here — no other machine on this page combines the two. It is also the one machine in this comparison that cannot give corosync its own port, so it does not appear in any cluster row.

The EQ14 in the two-node row is the cheap answer, not the comfortable one. Two ports satisfy the networking recommendation, which is the part most budget machines fail, but four threads and about 12GB of guest memory per node put a hard limit on what those nodes can host. On power, its 11.8W sits just under the MS-01’s 13.2W, but the first was measured under Windows and the second under Ubuntu, so read the two as roughly even. What puts the EQ14 in that cell is the lower purchase price, with two ports that are enough for corosync.

The row below it is where the EQ14 stops working, and the reason is Ceph rather than the node count. Three nodes ask no more of the network than two. Ceph asks for a dedicated 10Gbps link the EQ14 does not have, and its one OSD would take 4 to 8GB of the 14GB left after Proxmox. The MS-01 and MS-A2 fill that row because they have 10Gbps to give Ceph and memory to spare — not because they meet the guide’s 12-OSD layout, which none of these machines do.

The bottom row deserves a caveat. PCIe passthrough on consumer hardware depends on how the board groups devices under IOMMU, and that varies by firmware version in ways no specification sheet records. None of our sources tested passthrough on these machines. Treat the bottom row as “has a slot to pass through”, not as “passthrough is known to work”.

Frequently asked questions

Can I run Proxmox on a mini PC at all?

Yes. The requirement is a 64-bit CPU with virtualization extensions — Intel VT-x or AMD-V — and enough memory for the guests on top of the 2GB Proxmox asks for itself. All four machines here clear that. The constraints that bite are memory, NVMe slots and network ports, not CPU capability.

Do I need three machines to build a cluster?

For reliable quorum with High Availability, the Proxmox documentation says at least three nodes. Two nodes will form a cluster, but they cannot hold quorum when one is unavailable. The documented way around it is a QDevice supplying a third vote. It can run on hardware you already have, so two nodes plus a QDevice is a legitimate build, not a workaround.

Why does one network port matter if I only have one machine?

For a single node it does not. It matters when you add the second machine, because the documentation recommends a dedicated physical NIC for cluster traffic and a one-port machine has nothing to dedicate. If you will never cluster, the SER8’s single port costs you nothing.

Do I need ECC memory for ZFS?

The Proxmox ZFS documentation says “we recommend the use of high quality ECC RAM.” That is a recommendation, not a requirement. On these machines the choice is mostly made for you. For the MS-01’s i5-12600H, Intel ARK lists ECC Memory Supported: No. Neither Beelink nor Minisforum publishes ECC support for the other three either way. If ECC is a hard requirement, this is not the category to shop in.

How much does one of these cost to run?

The three-year column in the table above does the arithmetic per machine, and our general home server comparison works through the method in more detail.

Does a cluster need matching machines?

No. Proxmox will cluster machines with different CPUs and different amounts of memory, though the Ceph guide prefers identical servers. On mixing vendors the documentation is blunt: “Live migrations between Intel and AMD host CPUs have no guarantee to work.” Within one vendor it recommends setting the VM CPU type to the lowest generation in the cluster, which costs you the newer instructions on the faster nodes. Two of the machines here are Intel and two are AMD, so match the nodes that will share a workload, and let the odd one out be the QDevice.

Should I buy used enterprise hardware instead?

For Proxmox, the case for used enterprise gear is the rules this article is built around. A rack server has enough PCIe lanes and drive bays that the NVMe and OSD arithmetic stops mattering, enough network ports that corosync gets its own, and a platform that documents ECC support.

The trade is power and noise. A used tower or rack server typically idles somewhere between 40W and 100W, which at U.S. average rates is $64 to $160 a year, every year. That is a general range rather than a measurement we can cite, but it catches up with the purchase price faster than most people expect.

If you haven’t settled on Proxmox yet

This comparison assumes the hypervisor decision is made. If it is not, the specifications that matter shift — a machine that is wrong for a Proxmox cluster can be right for a single-purpose NAS or media box.