Running Ollama on a mini PC is the sweet spot for local AI. Not as cheap as a Raspberry Pi, not as expensive as a full desktop. A mini PC sits on your desk, uses 10-50W, and runs 7B-13B models at usable speeds.
Here are the best options in 2026.
Quick picks
| Budget | Pick | Price | Why |
|---|---|---|---|
| Best overall | Mac Mini M4 | $599 | Best performance/watt, unified memory |
| Best value | Beelink SER8 (AMD) | $350 | 32GB RAM, good CPU |
| Best budget | Beelink Mini S12 | $150 | Basic but works for 7B models |
| Best NPU | GEEKOM A9 Max | $700 | 55 TOPS NPU for AI tasks |
| Best RAM | Minisforum UM790 Pro | $450 | 64GB DDR5 option |
Mac Mini M4 ($599)
The best mini PC for Ollama, period.
| Spec | Value |
|---|---|
| CPU | Apple M4 (10-core) |
| GPU | 10-core integrated |
| NPU | 38 TOPS Neural Engine |
| RAM | 16GB unified (upgradeable to 32GB) |
| Storage | 256GB-2TB SSD |
| Power | 10-15W (idle to load) |
The M4’s unified memory architecture is perfect for LLMs. CPU, GPU, and NPU share the same memory pool, so there’s no copying overhead. 16GB runs 7B models well. 32GB runs 13B models.
Ollama performance:
| Model | Speed (16GB) | Speed (32GB) |
|---|---|---|
| Llama 3.1 8B Q4 | 18-22 tok/s | 20-25 tok/s |
| Qwen 3.6 7B Q4 | 15-20 tok/s | 18-22 tok/s |
| Gemma 4 9B Q4 | 12-16 tok/s | 15-18 tok/s |
| Llama 3.1 70B Q2 | N/A | 3-5 tok/s |
My take: If you can afford $599, get the Mac Mini M4. The performance, power efficiency, and software ecosystem are unmatched. The 16GB model handles 7B-8B models well. Upgrade to 32GB for 13B+ models.
Beelink SER8 ($350)
Best value for Ollama on a budget.
| Spec | Value |
|---|---|
| CPU | AMD Ryzen 7 8845HS |
| GPU | Radeon 780M (integrated) |
| RAM | 32GB DDR5 |
| Storage | 512GB NVMe |
| Power | 15-35W |
32GB DDR5 at $350 is excellent value. The Ryzen 7 8845HS has strong CPU performance and the Radeon 780M iGPU helps with inference.
Ollama performance:
| Model | Speed |
|---|---|
| Llama 3.1 8B Q4 | 10-14 tok/s |
| Qwen 3.6 7B Q4 | 8-12 tok/s |
| Gemma 4 9B Q4 | 7-10 tok/s |
My take: Best bang for buck. 32GB RAM handles 7B-8B models comfortably. Not as fast as the Mac Mini, but $250 cheaper.
Beelink Mini S12 ($150)
The cheapest option that actually works.
| Spec | Value |
|---|---|
| CPU | Intel N95/N100 |
| GPU | Intel UHD (integrated) |
| RAM | 16GB DDR5 |
| Storage | 500GB NVMe |
| Power | 10-15W |
Don’t expect miracles. The N95/N100 is a low-power CPU. But 16GB RAM and decent storage mean you can run 7B models at slow speeds.
Ollama performance:
| Model | Speed |
|---|---|
| Llama 3.1 8B Q4 | 3-5 tok/s |
| Qwen 3.6 7B Q4 | 2-4 tok/s |
| Gemma 4 2B Q4 | 8-12 tok/s |
My take: Only if budget is the primary constraint. 3-5 tok/s is usable for batch processing but painful for interactive chat. Spend the extra $200 for the Beelink SER8 if possible.
GEEKOM A9 Max ($700)
Best for NPU-accelerated AI.
| Spec | Value |
|---|---|
| CPU | AMD Ryzen AI 9 HX 370 |
| NPU | AMD XDNA 2 (55 TOPS) |
| RAM | Up to 64GB DDR5 |
| Storage | M.2 NVMe |
| Power | 15-35W |
55 TOPS NPU for AI tasks. The Ryzen AI 9 HX 370 is a powerful CPU with integrated Radeon 890M graphics.
My take: Only if you specifically need NPU acceleration. For Ollama, the CPU and RAM matter more than the NPU. The $350 Beelink SER8 is better value for pure LLM inference.
Minisforum UM790 Pro ($450)
Best for maximum RAM.
| Spec | Value |
|---|---|
| CPU | AMD Ryzen 9 7940HS |
| GPU | Radeon 780M (integrated) |
| RAM | Up to 64GB DDR5 |
| Storage | M.2 NVMe |
| Power | 15-35W |
64GB DDR5 option lets you run 13B-30B models. The Ryzen 9 7940HS is a strong CPU.
Ollama performance (with 64GB):
| Model | Speed |
|---|---|
| Llama 3.1 8B Q4 | 12-16 tok/s |
| Qwen 3.6 14B Q4 | 6-9 tok/s |
| Llama 3.1 70B Q2 | 2-3 tok/s |
My take: If you want to run larger models (14B+), the 64GB option is worth the extra $100 over the Beelink SER8.
What matters for Ollama
In order of importance:
- RAM: More RAM = larger models. 16GB minimum, 32GB recommended, 64GB for 14B+ models.
- CPU: Faster CPU = faster token generation. Ryzen 7+ or Apple M4+ recommended.
- Storage: NVMe SSD for fast model loading. 500GB minimum.
- GPU: Helps but not critical. Apple’s unified memory and AMD’s Radeon iGPU provide some acceleration.
- NPU: Nice to have but not essential for Ollama. Ollama doesn’t use NPUs directly.
My overall recommendation
$599 budget: Mac Mini M4. Best performance, best power efficiency, best software.
$350 budget: Beelink SER8. Best value, 32GB RAM, good performance.
$150 budget: Beelink Mini S12. Cheapest option, works for small models.
Need 64GB RAM: Minisforum UM790 Pro. Runs larger models.
For most developers, the Mac Mini M4 is the right choice. The unified memory architecture is ideal for LLM inference, and the performance/watt ratio is unmatched. If budget is tight, the Beelink SER8 at $350 is excellent value.
FAQ
How much RAM do I need for Ollama?
16GB minimum for 7B models. 32GB recommended for comfortable 7B-8B usage. 64GB for 14B+ models.
Is Mac Mini M4 good for Ollama?
Excellent. The unified memory architecture means CPU, GPU, and NPU share the same memory pool. No copying overhead. 16GB runs 7B models at 18-22 tok/s.
Can I run 70B models on a mini PC?
With 64GB RAM, you can run 70B models at Q2 quantization (2-3 tok/s). It’s slow but works for batch processing. Not recommended for interactive use.
Which mini PC has the best value?
Beelink SER8 at $350. 32GB DDR5, Ryzen 7 8845HS, good performance. Best bang for buck for Ollama.
Do I need a GPU for Ollama?
Not strictly. Ollama runs on CPU. But Apple’s unified memory and AMD’s Radeon iGPU provide some acceleration. A dedicated GPU is not necessary for a mini PC.