Skip to main content
PromptQuorumBuilt for humans. Structured for AI.

Best Mini PC for an Always-On Ollama Server (2026)

The Minisforum UM890 Pro (AMD Ryzen 9 8945HS, 32–96 GB RAM, Radeon 780M iGPU) is the best mini PC for an always-on Ollama server in 2026: runs 7B models at 12–18 tok/s CPU-only, draws 25–45 W under load, and fits inside a drawer.

Best Mini PC for an Always-On Ollama Server (2026)
Read in:
🇺🇸en🇩🇪de🇫🇷fr🇯🇵ja🇨🇳zh🇪🇸es🇧🇷pt🇸🇦ar🇰🇷ko

This page contains links to third-party products for reference. PromptQuorum is not enrolled in any affiliate program — these are plain links that earn no commission. Clicking links and your next steps are entirely your own responsibility. These links do not represent any endorsement or verification by PromptQuorum.

Hardware & PerformanceIntermediate

Key Takeaways

  • ✓Mini PCs draw 15–45 W vs 200–350 W for desktop GPUs — 24/7 savings matter
  • ✓UM890 Pro runs 7B models CPU-only at 12–18 tok/s; fine for API server use
  • ✓AOOSTAR GEM12 Pro + OCuLink eGPU unlocks GPU acceleration without a desktop PC
  • ✓Mac Mini M5 Pro: 64 GB unified memory is the ceiling for 30B-class models; base M6 tops out at 32 GB — both ship Sept 22, 2026
  • ✓Beelink SER8 is the <$400 starting point — 32 GB RAM for 7B and 13B

Best Mini PCs for Always-On Ollama Server — Ranked

📍 In One Sentence

The Minisforum UM890 Pro is the best mini PC for an always-on Ollama server because it combines fast CPU inference, up to 96 GB DDR5 RAM, and sub-50 W power draw.

💬 In Plain Terms

A mini PC running Ollama can serve AI responses to your whole home network around the clock, using less power than a lightbulb. More RAM = larger models. OCuLink = add a real GPU without a full desktop.

1

Minisforum UM890 Pro — Best Overall

The Minisforum UM890 Pro runs the AMD Ryzen 9 8945HS (8-core, up to 5.2 GHz) and supports up to 96 GB DDR5-5600 dual-channel RAM — enough to load Llama 3.3 70B at Q4 entirely in RAM. The Radeon 780M iGPU (12 RDNA3 CUs) accelerates 7B and 13B models at 8–14 tok/s via ROCm. CPU-only 7B Q4 speed: ~12–18 tok/s. Idle power: ~15 W. Load (GPU active): ~35–45 W. Price: ~$500 (32 GB/1 TB) to $600 (64 GB).

Minisforum UM890 Pro on Amazonproduct link · disclosedMinisforum UM890 Pro on Minisforum.comproduct link · disclosed
2

AOOSTAR GEM12 Pro OCuLink — Best for eGPU

The AOOSTAR GEM12 Pro OCuLink adds an OCuLink port that connects to an external GPU at PCIe 4.0 x4 bandwidth (64 Gbps) — enough to run an RTX 3090 at full speed for Ollama. Without eGPU: same as other mini PCs, 13–18 tok/s CPU-only. With RTX 3090 eGPU: 60–80 tok/s on 7B Q4. The mini PC itself runs AMD Ryzen 7 Pro 8845HS with 32–96 GB DDR5. OCuLink to PCIe adapter required (~$30). Barebone price: ~$380; configured (32 GB/1 TB): ~$480.

AOOSTAR GEM12 Pro OCuLink on Amazonproduct link · disclosed
3

Beelink SER8 — Best Budget Pick

The Beelink SER8 runs the AMD Ryzen 7 8745HS (8-core, up to 4.9 GHz) with 32 GB DDR5 and 1 TB NVMe SSD for ~$450. CPU-only Ollama speed: ~10–15 tok/s on 7B Q4. Idle power: 10–15 W. RAM is user-upgradeable (not soldered on current 8745HS revision). If you want a solid entry into always-on Ollama without spending $500+, the SER8 covers 7B and 13B models well.

Beelink SER8 on Amazonproduct link · disclosed
4

Apple Mac Mini M6 / M5 Pro — Best for macOS

Apple refreshed the Mac Mini lineup on August 25, 2026, replacing the M4 generation. The Mac Mini M6 starts at $899 with up to 32 GB unified memory and 170 GB/s bandwidth. The Mac Mini M5 Pro starts at $1,699 with up to 64 GB unified memory, 307 GB/s bandwidth, and Thunderbolt 5. Both ship September 22, 2026. For an always-on Ollama server, the memory ceiling is what matters: 32 GB (M6) comfortably fits 7B–13B models with headroom for the OS, while 64 GB (M5 Pro) is what you need to load 30B-class models at Q4 without spilling to swap. No independent Ollama benchmarks exist yet for either chip — PromptQuorum has not tested this hardware, and tok/s figures will follow once real-world numbers are available.

Apple Mac Mini M6 on Amazonproduct link · disclosedApple Mac Mini on Apple.com (M6 and M5 Pro configs)product link · disclosed

Always-On Electricity Cost Comparison

At $0.15/kWh (US average), running 24/7 for 30 days:

DeviceAvg Load PowerMonthly Cost (24/7)
Minisforum UM890 Pro35 W~$3.78/mo
Beelink SER825 W~$2.70/mo
Mac Mini M6 / M5 ProNot yet benchmarkedTBD after independent tests
Desktop RTX 4060 Ti (comparison)200 W~$21.60/mo
Cloud API (GPT-5.5-mini, 1M tok/day)N/A~$45–90/mo

Related Guides

Quick Answers

Can a mini PC run 13B or larger models at useful speed?▾
Yes — with enough RAM. The Minisforum UM890 Pro with 64 GB runs Llama 3.3 13B Q8 entirely in RAM at ~8–12 tok/s CPU-only. With the Radeon 780M iGPU accelerating, Q4 models run at 10–18 tok/s — usable for background summarization or API calls. Interactive chat benefits from at least 12–15 tok/s. For 30B+ models, the Mac Mini M5 Pro (64 GB unified memory, from $1,699) has enough headroom — that pushes it above the $1,500 mark. The base Mac Mini M6 ($899, 32 GB max) does not have enough memory for 30B-class models at a useful quant level.
Does Ollama work well as a network server on a mini PC?▾
Yes. Set OLLAMA_HOST=0.0.0.0 in your environment and Ollama serves requests from any device on your LAN. Pair with Open WebUI (Docker container) for a browser-based interface accessible from phones, tablets, and PCs. The mini PC draws low power, runs silently, and handles one concurrent request at a time without issue.
What about eGPU setups — are they worth it?▾
For Ollama specifically, an OCuLink eGPU (AOOSTAR GEM12 Pro + RTX 3090 enclosure) is the best of both worlds: desktop GPU speed with mini PC power draw when idle. OCuLink (PCIe 4.0 x4) delivers ~80% of the bandwidth of a direct PCIe x16 slot — enough for LLM inference with minimal bottleneck. Thunderbolt eGPUs are slower (~40% bandwidth) and not recommended for GPU-intensive inference.

Want the full breakdown?

Read the complete guide →
← Back to Prompt Bites