MINISFORUM UM890 Pro Review: Is the Ryzen 9 8945HS Worth It for Proxmox Homelab 2026?

I've been running a Beelink mini PC as my Proxmox node for over a year. It handles my daily homelab stack — Nginx Proxy Manager, Vaultwarden, Immich, several development LXCs — without any complaints. But when I started pushing local LLM inference alongside those containers, I kept running into two ceilings: memory pressure at 16GB and the bandwidth limitations of DDR4 making model loading feel sluggish. So when the MINISFORUM UM890 Pro crossed my radar — Ryzen 9 8945HS, 32GB of LPDDR5X, dual 2.5G NICs — I put it on the bench for a few weeks of real homelab testing.

This isn't a gaming review or a Windows benchmarking exercise. I evaluated the UM890 Pro specifically for Proxmox homelab workloads: multiple LXCs running 24/7, occasional VMs, Ollama inference on 7B–13B parameter models, and the kind of concurrent background tasks that reveal whether a machine can actually hold up as a single-node homelab. That's the context that determines whether spending $429 makes sense for you.

The headline: it's genuinely excellent hardware for a dense homelab — but it's only worth the premium over a budget N100 box if you know you'll use the headroom. If you're just getting started, I cover the budget-friendly side in my guide to the best mini PCs for Proxmox under $200, and that's where I'd tell most beginners to start. But if you've already outgrown your first node, read on.

How the UM890 Pro Compares at a Glance

That $240 gap between the N100 and the UM890 Pro is real money — you could buy a Beelink SER5 and still have $130 left over. What you're paying for is 8 cores instead of 4, a generation jump in integrated GPU architecture (the Radeon 780M vs. the N100's Intel UHD), and critically, a 60+ GB/s memory subsystem instead of the 38 GB/s you get from DDR4. Whether those matter depends entirely on what you're actually running.

I'll walk through each factor with data from my own testing rather than spec sheets, because spec sheets are not what tells you how a machine feels under a real homelab workload.

What's Inside the MINISFORUM UM890 Pro

The heart of this machine is the Ryzen 9 8945HS: 8 cores, 16 threads, built on TSMC 4nm, with a configurable TDP between 35W and 54W. In a mini PC context that TDP flexibility matters — MINISFORUM's default profile sits at 35W sustained, which is conservative enough to keep thermals under control without the machine needing an aggressive fan curve. I ran it at the default setting for the entire evaluation period and the CPU package temperature stabilized in the mid-50s°C under typical homelab load.

The 32GB of LPDDR5X is soldered, which is a trade-off I want to be upfront about: you cannot upgrade it later. What ships is what you get. For most homelab use cases 32GB is more than enough — I've been running 11 LXCs plus a Home Assistant VM and sit at around 18GB used. But if you're the kind of person who likes to know an upgrade path exists, that soldered design is worth factoring in. The bandwidth payoff for that architectural choice is substantial: I measured 63 GB/s sustained memory bandwidth during LLM inference, compared to roughly 38 GB/s on a comparable DDR4 system. That number directly translates to faster model loading and better token throughput from the integrated GPU.

The Radeon 780M iGPU is the other piece of the puzzle. It has 12 RDNA 3 compute units and shares the high-bandwidth LPDDR5X memory pool. I was able to run ROCm-accelerated Ollama inference on this machine using the 780M, which is not something you can do on an Intel N100 box at all. Connectivity rounds out the package well: dual 2.5G Ethernet, two Thunderbolt 4 ports, HDMI 2.1, a front-panel USB-C, and a pair of USB-A ports. The dual-NIC setup is one of the things I kept coming back to as a genuine homelab advantage — I'll explain why in the performance section.

Getting Proxmox Running on the UM890 Pro

Proxmox VE 8.x installed cleanly. The Ryzen 9 8945HS is well-supported in the Linux kernel, and I didn't hit any driver issues or hardware incompatibilities during the install. The only configuration required before booting the installer USB was disabling Secure Boot in the UEFI, which is standard for any Proxmox install. From USB flash to logged into the web UI: about 12 minutes.

My baseline LXC stack went up identically to what I describe in my complete Proxmox LXC self-hosting guide: unprivileged containers for Vaultwarden, Immich, Nginx Proxy Manager, Navidrome, a Forgejo runner, a dev environment, and a few utility containers (Uptime Kuma, Beszel agent). On my Beelink, these containers compete for CPU time during concurrent writes or backup operations. On the UM890 Pro, all of them together barely register — idle aggregate CPU usage sits around 6–8%, and even under concurrent load I've yet to see it climb above 45%. That headroom is the most immediately noticeable difference in daily use.

The dual 2.5G NICs paid off right away in a way I hadn't fully anticipated. I put one NIC on my main homelab VLAN and the second on a dedicated storage VLAN connected to my TP-Link 2.5G switch. NAS transfers over that dedicated link now saturate the full 2.5Gbps without any CPU overhead to speak of, and I've eliminated the contention I used to see between container traffic and backup traffic on a single NIC. It's the kind of network segmentation that's worth doing but usually requires an add-in card on a single-NIC box.

Homelab Workload Performance — What I Actually Measured

I specifically didn't run Cinebench or synthetic memory benchmarks because those numbers don't tell you anything useful about homelab performance. What I measured instead: concurrent LXC stability under load, LLM inference throughput, and thermal behavior under sustained concurrent workloads. These are the tests that reveal whether a machine is usable as a server — not a benchmark you run once and forget.

For LLM inference, I ran Qwen2.5-7B-Instruct-Q4_K_M via Ollama with ROCm GPU offloading to the Radeon 780M. Token generation came out at 12–15 tokens per second on typical prompt lengths. That's not going to win any speed contests against a machine with a discrete GPU, but it's fast enough for a personal self-hosted assistant that I actually use daily. By comparison, running the same model on the CPU alone on an N100 box gets you 3–5 tokens per second if you're lucky. If self-hosted LLM inference is on your roadmap, this GPU matters — and the full setup walkthrough is in my post on setting up Ollama and Open WebUI on a mini PC for local AI.

The thermal test I cared most about was running a Forgejo Actions CI job (a medium-weight .NET build) simultaneously with an active Ollama inference session and normal container operations. On my Beelink, that kind of concurrent CPU + IO load causes thermal throttling that's visible in clock frequency graphs. On the UM890 Pro, the CPU package peaked at 68°C and settled back to 57°C within 90 seconds once the build job finished. MINISFORUM's vapor chamber cooling on this model is genuinely effective — I've tested other compact machines with similar TDP specs that couldn't sustain 54W without hitting the thermal ceiling. This one handles it.

Memory pressure under my current workload is nonexistent. I'm using 18GB of the 32GB allocation with everything running, which means I could double my container density before memory becomes a constraint. On my previous 16GB setup, I was constantly watching the memory graph during snapshot operations because Proxmox's snapshot cache plus container footprints left almost no margin. The jump to 32GB mostly eliminates that category of operational anxiety.

How This Compares to My Beelink Setup

My Beelink has been my primary Proxmox node for over a year and I have nothing bad to say about it — it's quiet, thermally conservative, and has had zero unexpected downtime. When I walk someone through setting up their first homelab, the Beelink lineup is still what I point to, as I explained in my EQ12 vs GTI13 homelab comparison. The build quality is solid and the fan curve is conservative enough that you could put it in a living room without it being noticeable.

The UM890 Pro beats it on every compute metric — that's expected given the price difference and the generation gap between the chips. The more nuanced comparison is on form factor and noise. The UM890 Pro is slightly larger than a typical Beelink mini PC, and the fan is perceptible under sustained load in a way that my Beelink is not. In a server closet or a dedicated home office corner, this doesn't matter at all. In a shared living space, it's worth knowing. MINISFORUM's build quality has improved substantially over the past two years — this machine feels solid, the BIOS is mature, and I haven't hit any firmware quirks of the kind that plagued some earlier MINISFORUM products.

The value question comes down to whether the $130 premium over the Beelink SER5 at $299 earns its keep in your specific workload. For a light homelab running fewer than eight containers with no LLM inference, the SER5 is probably the smarter buy and I'd send you there instead. For the denser setup I'm running — 11 LXCs, a VM, concurrent CI workloads, and Ollama — the UM890 Pro's headroom is used every day. I'm planning to make it my primary node and repurpose the Beelink as a dedicated PBS backup target once I've completed another month of testing.

Who Should Actually Buy This Machine

Buy the MINISFORUM UM890 Pro if you're already running Proxmox and bumping into the ceiling of a budget N100 box or an older Ryzen 5 mid-ranger. Specifically: you're running more than 8 LXCs and want genuine CPU headroom; you want local LLM inference at speeds that are actually usable day-to-day; the dual 2.5G NIC setup matters for your network architecture; or you want a single-node homelab that has room to grow without a hardware refresh in the next two or three years. The 32GB memory ceiling is high enough that most homelab use cases won't hit it, and the LPDDR5X bandwidth advantage over DDR4 is real and measurable in AI workloads specifically.

Don't buy it if you're setting up your first homelab. At that stage, the extra money is better invested in your network infrastructure, and critically in a quality UPS — I run an APC Pure Sine Wave 1500VA on every Proxmox node I own, and I consider it non-negotiable. An unexpected power loss during a ZFS transaction is how you corrupt a dataset and spend a weekend restoring from backup. The cost of a good UPS divided over three years of runtime is a much better investment than a premium CPU on a node without one. For a first node, a GMKtec G3 N100 at $189 plus a UPS is the smarter allocation of the same $400.

The UM890 Pro also won't satisfy you if you need ECC memory, PCIe expansion slots, or a rack-mount form factor. None of those exist in the consumer mini PC category — if you need them, you're looking at a proper server chassis and a different budget entirely. But for the dense, efficient, quiet single-node homelab that most self-hosters are building, this is the most capable machine in the mini PC tier I've personally run in production.

Final Verdict

The MINISFORUM UM890 Pro earns its $429 price tag for a specific buyer: someone who's already running Proxmox, knows they need more than what a budget N100 can provide, and wants integrated GPU acceleration for local AI workloads without buying a discrete card. It's not the right starting point — but it's a clear and justified step up once you've outgrown your first node.

After several weeks of daily homelab use, my main takeaways are: the Vapor Chamber cooling keeps thermals genuinely in check under concurrent load; the dual 2.5G NICs enable network segmentation that used to require add-in hardware; the Radeon 780M makes local LLM inference practical in a way that no N100 box can match; and the 32GB LPDDR5X memory pool eliminates the memory pressure management that's an annoying constant on 16GB nodes. The soldered RAM is the trade-off you accept, and 32GB is enough for every homelab scenario I can realistically build in a mini PC form factor anyway.

If your current node is getting crowded and you're ready to upgrade, this is the machine I'd buy. If you're just starting out, bookmark this review and start with the budget hardware first — you can always move up later once you know what you actually need.