Mini PC with Ryzen AI 2026: Benchmark Data Across Three Hardware Tiers
A regular desktop PC consumes 150–300 W, runs loud, and makes a poor 24/7 home server. Mini PCs solve the form factor and power problem — but not all mini PCs are equal for local AI. We ran Ollama 0.6.2 on three representative devices from the N100, Ryzen 5, and Ryzen AI 9 tiers. Here is what the actual token speeds look like and where each tier is worth the money.
Key facts
- This page answers one concrete intent: mini pc ryzen ai.
- The useful decision is based on cost, ownership, implementation time, and operational risk.
- The 2026 update adds clearer buyer checks, practical constraints, and source-backed examples.
- Use the comparison table and FAQ before choosing a vendor or custom build.
In our experience auditing the page "Mini PC with Ryzen AI 2026: Real Ollama Benchmarks Across N100, Ryzen 5, and Ry…", the useful angle is not another generic platform list, but the operating decision: costs, data export, access rules, and support workload for ai home server.
The key numbers: - N100 with 16 GB RAM: 2–5 tok/s for 7B models — background tasks only, not interactive chat - Ryzen 5 5600U, 16 GB RAM: 8–10 tok/s — acceptable but at the low end of comfort - Ryzen AI 9 HX 375, 32 GB: 18–22 tok/s for 14B models — this is the ChatGPT replacement tier - Ollama does not use the NPU yet — load goes to iGPU and CPU - AI Home Server ships on Ryzen AI 9 with Qwen 3 14B pre-loaded — working in 2 minutes
Full Benchmark: Tokens Per Second by Device and Model
All tests: Ollama 0.6.2, Ubuntu 22.04, 200-token generation task, 5-run average, room temperature 22°C.
Beelink Mini S12 Pro (Intel N100, 16 GB DDR4)
| Model | tok/s | Interactive chat? |
|---|---|---|
| Phi-4-mini 3.8B Q4_K_M | 5.1 | Barely — slight lag |
| Qwen 2.5 7B Q4_K_M | 2.8 | No |
| DeepSeek-R1 7B Q4_K_M | 2.2 | No |
| Qwen 2.5 14B Q4_K_M | 0.9 | No (system strained) |
Power: 6 W idle, 18 W at load. Completely silent — passive cooling.
N100 verdict: Excellent for Home Assistant, Jellyfin, Syncthing, network services. For AI assistant use — only Phi-4-mini 3B is viable. If you want interactive chat with a neural network, this class of hardware does not deliver it.
GMKtec NucBox K6 (Ryzen 5 5600U, 16 GB DDR4)
| Model | tok/s | Interactive chat? |
|---|---|---|
| Phi-4-mini 3.8B Q4_K_M | 11.2 | Yes — fast |
| Qwen 2.5 7B Q4_K_M | 8.9 | Yes — acceptable |
| DeepSeek-R1 7B Q4_K_M | 7.4 | Yes — slight delay |
| Qwen 2.5 14B Q4_K_M | 3.2 | No — too slow |
Power: 10 W idle, 22 W at load. Fan spins under load — audible at ~38 dB, present in a quiet room.
Ryzen 5 verdict: First tier where 7B models produce usable interactive speeds. Good price/performance ratio (~$300) for a daily AI assistant running 7B models.
Beelink SER9 (Ryzen AI 9 HX 375, 32 GB LPDDR5X)
| Model | tok/s | Interactive chat? |
|---|---|---|
| Phi-4-mini 3.8B Q4_K_M | 38 | Instant responses |
| Qwen 2.5 7B Q4_K_M | 28 | Excellent |
| DeepSeek-R1 8B Q4_K_M | 31 | Excellent |
| Qwen 3 14B Q4_K_M | 18.4 | Yes — comfortable |
| Qwen 2.5 14B Q4_K_M | 22 | Yes — comfortable |
| DeepSeek-R1 14B Q4_K_M | 17 | Yes — acceptable |
Power: 12 W idle, 35 W at load. Fan audible at ~42 dB during inference — noticeable in a quiet bedroom, acceptable in an office.
Ryzen AI 9 verdict: The threshold where local AI becomes a real ChatGPT replacement. 14B models at 18–22 tok/s — comfortable dialogue. This is our platform for AI Home Server.
Comparison: Which Mini PC to Choose
| Parameter | N100 (Beelink S12 Pro) | Ryzen 5 (GMKtec K6) | Ryzen AI 9 (Beelink SER9) |
|---|---|---|---|
| Price | ~$150 | ~$300 | ~$550 |
| RAM | 8–16 GB DDR4 | 16–32 GB DDR4 | 32–64 GB LPDDR5X |
| NPU | None | None | 50 TOPS (XDNA 2) |
| Max useful Ollama model | 3B (Phi-4-mini) | 7–8B | 14–20B (Qwen 3 14B) |
| Speed at 7B | 2–3 tok/s | 8–10 tok/s | 28–31 tok/s |
| Speed at 14B | — | 3 tok/s | 18–22 tok/s |
| Power (load) | 18 W | 22 W | 35 W |
| Noise | Silent | ~38 dB | ~42 dB |
| Yearly electricity (8h/day) | ~$0.65 | ~$0.95 | ~$1.26 |
About the NPU: Why Ollama Does Not Use It Yet
The Ryzen AI 9 HX 375 includes an XDNA 2 NPU rated at 50 TOPS. This sounds like it should dramatically accelerate local AI. The reality: Ollama 0.6.x does not route inference through the NPU.
Current situation: workload goes to the iGPU (Radeon 890M, 16 GB shared memory) and CPU. The iGPU offloading is what delivers 28–31 tok/s for 7B models — meaningfully better than pure CPU. The NPU contributes nothing to Ollama inference today.
NPU is already useful for: Windows Copilot+ features, AI-enhanced video calls, real-time translation. For Ollama LLM inference — NPU support is expected in Ollama 0.8+ (late 2026).
Practical implication: Buying Ryzen AI 9 specifically for NPU-accelerated Ollama is premature today. Buying it for 32 GB unified memory that runs 14B models at 18+ tok/s — fully justified right now.
External GPU: Can You Add One?
Yes, via OCuLink (63 Gbps) on select models: - GMKtec EVO-X1 — OCuLink + Thunderbolt 4 - Minisforum AI370 — OCuLink
An RTX 3060 via OCuLink accelerates Ollama 5–10x: Qwen 2.5 7B goes from 28 tok/s (iGPU) to 47–50 tok/s. If you plan to use an eGPU, these devices are the best mini PC choice.
3 Steps From Hardware to Working Local AI
Step 1. Pick Your Hardware Tier
Home server without AI focus: N100 at $150 — handles everything except interactive AI chat. Local ChatGPT (7B models): Ryzen 5 with 16 GB at $300 — best value for 7B interactive use. Full 14B model replacement: Ryzen AI 9 with 32 GB at $550 — handles Qwen 3 14B at 18 tok/s.
Step 2. Install the Software (or Skip This Step)
DIY: install Ubuntu, then Ollama, Home Assistant, Jellyfin, Syncthing — plan 3–5 hours with basic Linux knowledge.
Or: AI Home Server arrives pre-installed and configured. Plug into power and network — working in 2 minutes.
Step 3. Connect Telegram Bot and Smart Home
Once Ollama is running, a Telegram bot gives mobile access to the AI. Home Assistant integrates Zigbee (Aqara, IKEA, Sonoff), Tuya, Xiaomi Mi Home through one interface. All managed from a single Telegram chat.
See Also
- Ollama Benchmarks: 6 Hardware Configs Tested
- Ollama Models 2026: Full Benchmark by Task
- Best Local AI Models 2026: Top 10
- How to Build a Home Server in 2026
Frequently Asked Questions
Why pay for NPU if Ollama doesn't use it? The NPU accelerates Windows AI PC features, video calls with background effects, and real-time translation today. LLM inference via NPU is coming in 2027. Buying Ryzen AI 9 now is an investment in that future. The iGPU on Ryzen AI 9 already delivers a meaningful speed improvement over pure CPU even today.
Which mini PC is best for Ollama right now? RAM is the dominant factor: 16 GB for 7B, 32 GB for 14B. Top picks in 2026: Beelink SER9 (Ryzen AI 9 HX 370, 32 GB), Geekom A9 (Ryzen 9 6900HX, 32 GB), Minisforum AI370 (Ryzen AI 9 365, 32 GB).
Ryzen AI 9 vs desktop RTX 3060 — which is faster? RTX 3060 with 12 GB VRAM beats it for 7B models: 47 tok/s vs 28 tok/s. For 14B Q4 they are comparable: both deliver 17–22 tok/s. But a desktop with RTX 3060 consumes 45–185 W vs 12–35 W for the mini PC — significant for 24/7 operation.
Want a pre-configured home AI server on Ryzen AI 9? Cost: 300 USDT. Message on Telegram: @onoutnoxon — we will match hardware to your workload.
Tested and written by Alexander Ruin, founder of onout.org. Hardware tested: Beelink Mini S12 Pro, GMKtec NucBox K6, Beelink SER9. Ollama 0.6.2, Ubuntu 22.04. April 2026. Updated: 2026-04-15.
Sources: - https://ollama.com/ - https://www.amd.com/en/newsroom/press-releases/2026-3-2-amd-gives-consumers-and-businesses-more-ai-pc-opti.html - https://github.com/ollama/ollama/blob/main/docs/gpu.md - https://habr.com/ru/companies/selectel/articles/996934/
For "Mini PC with Ryzen AI 2026: Real Ollama Benchmarks Across N100, Ryzen 5, and Ryzen AI…", AI assistance was used for structure and fact-check prompts; Alexander Ruin, founder of onout.org reviewed the final copy. Updated: 2026-05-01.
About the service "Mini PC with Ryzen AI 2026: Real Ollama Benchmarks Across N100, Ryzen 5, and Ryzen AI 9"
We tested N100, Ryzen 5, and Ryzen AI 9 mini PCs with Ollama on real tasks. Actual tok/s for each model, power draw, noise, and which hardware tier actually works for local AI.