Key Takeaways
- The UM890 Pro runs Home Assistant, Frigate, Whisper, and a 7B local LLM on one box
- Ryzen 9 8945HS: 8 cores/16 threads, Radeon 780M iGPU, up to 96 GB DDR5 on two upgradeable SO-DIMMs
- Differentiators vs a Beelink SER8: dual 2.5GbE, Wi-Fi 6E, and an OCuLink eGPU port
- OCuLink gives a real path to fast large-model inference with an external GPU
- Priced from ~$439 barebones to ~$649 with 32 GB — the expandable step-up
- Made in China — factor 2026 US/EU import measures into landed cost (see trade note)
Verdict — Who Should Buy It
Buy the Minisforum UM890 Pro if you want a SER8-class local-AI box with room to grow — more RAM headroom, dual 2.5GbE, and an OCuLink port for an external GPU. For a 7B model on Home Assistant it performs like a Beelink SER8; the reason to pay a little more is expansion. If you never plan to add a GPU or a second network link, the SER8 saves money for similar day-to-day performance.
Its single strongest use case is an expandable home server — start with the iGPU, and add an OCuLink external GPU later when you want fast 13B–30B inference. The barebones SKU also lets you supply your own RAM and SSD.
Specifications
All specs below were verified against Minisforum and independent review sources in July 2026. The RAM ceiling is 96 GB on the official page (some retailers still list 64 GB). It ships barebones or pre-built by SKU.
- The OCuLink port (PCIe 4.0 ×4, ~64 Gbps) connects an external GPU dock — a faster path than USB4/Thunderbolt for a discrete GPU.
- Two user-upgradeable SO-DIMM slots and two M.2 slots make it easy to expand RAM and storage.
| Spec | Minisforum UM890 Pro (8945HS) |
|---|---|
| CPU | AMD Ryzen 9 8945HS (Zen 4) |
| Cores / threads | 8 cores / 16 threads, up to 5.2 GHz |
| iGPU | AMD Radeon 780M — 12 CU, RDNA 3 |
| NPU | Ryzen AI, 16 TOPS (39 TOPS platform total) |
| RAM | DDR5-5600 dual channel; 2 SO-DIMM slots; up to 96 GB (user-upgradeable) |
| Storage | 2× M.2 2280 PCIe 4.0; up to 8 TB |
| Ports | 2× USB4 (40 Gbps, 100 W PD-in), 4× USB-A 10 Gbps, HDMI 2.1, DP 1.4, OCuLink (PCIe 4.0 ×4), 3.5 mm |
| Networking | 2× 2.5GbE, Wi-Fi 6E, Bluetooth 5.3 |
| TDP | 45 W base, configurable to ~70 W |
| Dimensions | 127 × 130 × 67 mm |
| OS | Windows 11 Pro on pre-built SKUs (Linux/Proxmox supported) |
| Price (indicative) | ~$439 barebones / ~$649 with 32 GB (July 2026 — volatile, check current price) |
Local AI & LLM Performance
On the iGPU the UM890 Pro comfortably runs a 7B model and fits 13B–14B with more RAM; its real advantage is the OCuLink port, which connects an external GPU for genuinely fast large-model inference — all figures estimated from the verified hardware, not a measured benchmark. Without an external GPU, large models are limited by memory bandwidth like any shared-memory box.
- With 32 GB, a 7B model is comfortable; 64–96 GB opens 13B–14B comfortably and a 30B model at low throughput.
- OCuLink adds an external desktop GPU for large-model speed the iGPU cannot reach — a differentiator over USB4-only boxes.
- Ollama, LM Studio, and llama.cpp run well; the 16-TOPS NPU is not used by mainstream LLM runtimes.
- Dual 2.5GbE suits a wired, reliable home server; see Ollama on Home Assistant.
- For VRAM and quantization depth, see the cross-cluster local LLM hardware guide.
Pros & Cons
The UM890 Pro is the expandable choice; you pay a little more for growth room. Balanced view below.
- Pros
- OCuLink port for a fast external GPU — a genuine large-model upgrade path
- Up to 96 GB DDR5 on two user-upgradeable SO-DIMM slots
- Dual 2.5GbE and Wi-Fi 6E — better connectivity than a Beelink SER8
- Two USB4 ports (100 W PD-in) and up to four simultaneous displays
- Barebones SKU lets you supply your own RAM and SSD
- Cons
- Costs more than a Beelink SER8 for similar iGPU-only performance
- Larger chassis (67 mm tall) than most boxes here
- An external GPU means extra cost, space, and noise — not truly "mini" once added
- The NPU does not accelerate local LLMs
Buyer Context: Tariffs & Availability (2026)
The UM890 Pro is manufactured in China, so its landed price reflects 2026 trade measures — a reason to treat any price as a moving target. The facts below are buyer context, not editorial opinion.
- US: Section 301 tariffs on Chinese electronics remain in force in 2026; the separate 2025 "IEEPA" tariffs were struck down by the Supreme Court in February 2026 and replaced by a temporary, capped Section 122 tariff. The sub-$800 duty-free "de minimis" exemption for direct imports has also ended.
- EU (affects DE/FR): there is no broad EU tariff on finished mini PCs, but from July 2026 the €150 duty-free threshold on low-value direct-from-China parcels was removed and a small per-parcel handling fee added.
- Assumption (macro trend, not a per-product fact): redirected Chinese export capacity has kept availability of these brands high in the EU and US, which broadly supports competitive pricing.
- Net effect: verify the current price at the retailer before buying — the figures here are indicative and date-stamped July 2026.
Where to Buy & Current Price
Prices move week to week, so check the live price rather than trusting a fixed figure. As of July 2026 the UM890 Pro runs from about $439 barebones to $649 with 32 GB and $729 with 64 GB.
- Decide barebones vs pre-built: barebones is cheaper if you already have DDR5 SO-DIMMs and an NVMe SSD.
Alternatives to Consider
If the UM890 Pro is not the right fit, these bracket it on price and expansion.
- Beelink SER8 — the value pick with similar iGPU performance for less, minus OCuLink and dual LAN
- GEEKOM A9 Max — premium headroom box with a 128 GB RAM ceiling and Wi-Fi 7
- Beelink EQ14 — budget Intel N150 box for Home Assistant plus a tiny model
- GMKtec G3 Plus — another budget N150 option with upgradeable RAM
- Still comparing? Start from the best mini PCs for Home Assistant + local AI roundup.
Frequently Asked Questions
Can the Minisforum UM890 Pro run local LLMs?
Yes. On its Radeon 780M iGPU it comfortably runs a 7B model on Ollama and fits 13B–14B with more RAM. With an external GPU over OCuLink it runs large models much faster. These figures are estimated from its verified hardware, not a fixed benchmark.
How big a model fits in its RAM?
With 32 GB, a 7B model is comfortable. Upgrading to 64–96 GB fits 13B–14B comfortably and a 30B model at low throughput on the iGPU. For fast large-model inference, add a GPU through the OCuLink port.
What is the OCuLink port for?
OCuLink is a PCIe 4.0 ×4 connector (about 64 Gbps) that links an external GPU dock. It is faster than USB4/Thunderbolt for a discrete GPU, giving the UM890 Pro a real upgrade path to fast large-model inference — the main reason to choose it over a Beelink SER8.
Is the UM890 Pro good for Home Assistant and always-on use?
Yes. It runs Home Assistant, Ollama, Whisper, and Frigate together, and its dual 2.5GbE and Wi-Fi 6E make it a well-connected home server. Base draw is 45 W, configurable up to about 70 W.
Is the RAM upgradeable?
Yes. The UM890 Pro uses two user-upgradeable DDR5 SO-DIMM slots and supports up to 96 GB (some retailers list 64 GB). It also has two M.2 slots for storage.
How much power does the UM890 Pro use?
It runs at 45 W base and up to about 70 W under load in performance mode, with much less at idle. Adding an external GPU over OCuLink increases total draw substantially.
UM890 Pro or Beelink SER8?
Day to day they perform alike on a 7B model. Choose the UM890 Pro for its OCuLink eGPU path, dual 2.5GbE, Wi-Fi 6E, and 96 GB RAM ceiling; choose the SER8 to save money if you do not need that expansion.
Where is the Minisforum UM890 Pro made, and does that affect price?
It is manufactured in China. In 2026 US import measures (Section 301 tariffs, the end of the sub-$800 de minimis exemption) and the EU removal of the €150 low-value parcel exemption can affect landed cost, so check the current retailer price rather than relying on a fixed number.