Best Mini PC for Ollama in 2026: Specs That Decide It

The spec that decides how fast Ollama runs on a mini PC is memory bandwidth, not the CPU badge. A mini PC has no dedicated VRAM — the integrated GPU reads the model over the same DDR5 bus as the CPU — so the box to buy is a Ryzen 7 8845HS or Ryzen 9 8945HS (Radeon 780M, 12 RDNA 3 compute units) with 2 × 16 GB of dual-channel DDR5-5600, typically $520-$800. That configuration works out to roughly 90 GB/s, which puts the arithmetic ceiling for an 8B model at Q4_K_M near 18 tokens/sec. Buy the same chip with a single RAM stick and you have halved that ceiling before running anything.

Below: the arithmetic that predicts the token rate on any box, a spec table for the five mini PCs people actually shortlist, and the buying traps that cost more performance than a CPU upgrade will ever win back.

What actually decides token speed on a mini PC?

One number, and it is not the one on the marketing page.

Generating a token means reading the model's weights out of memory — every token, every time. So the ceiling is:

tokens/sec ceiling = memory bandwidth (GB/s) ÷ model size in memory (GB)

Memory bandwidth on a mini PC is itself arithmetic, not a mystery spec:

bandwidth (GB/s) = memory speed (MT/s) × bus width (bits) ÷ 8

A DDR5 SODIMM is 64 bits wide. Two sticks give a 128-bit bus; one stick gives 64 bits. That is the whole reason "dual-channel" keeps coming up in local-AI threads.

Model sizes come from the rule of thumb used across our hardware guides: Q4_K_M costs about 0.6 GB per billion parameters. That puts Llama 3.1 8B at roughly 4.8 GB and Phi-3 Mini 3.8B at roughly 2.3 GB — and the published Ollama download for llama3.1:8b-instruct-q4_K_M is 4.9 GB, so the rule checks out against the real artifact. Run both numbers through the formula and you get this:

Every figure there is a ceiling, not a forecast. Real output sits below it: the integrated GPU never saturates the bus, the OS and CPU are reading the same memory, and the KV cache grows as your context does. What a ceiling is good for is ranking decisions before you spend anything — and it says, bluntly, that the second RAM stick buys more than the faster CPU, and that a used discrete GPU has roughly four times the headroom of any mini PC on this list.

If the quantization side of that is new to you, quantization explained covers why Q4_K_M is the default, and our Ollama model RAM and VRAM table lists sizes model by model. The memory bandwidth explainer goes deeper on why this one spec dominates everything else.

Which mini PCs are worth shortlisting?

These five come up in every "mini PC for Ollama" thread. The striking thing about the table is what it shows: four of the five have identical memory bandwidth, so the CPU model is not what separates them.

Prices move constantly and SKUs vary by region, so treat them as a bracket. More importantly, confirm the memory configuration on the exact listing: vendors ship the same chassis with one stick or two, and the product title rarely tells you which. AMD's Ryzen 8000 series spec page is the authority for compute-unit counts and TDP envelopes on the AMD trio.

So what does separate them?

Not the ceiling — everything except the N100 shares it. Three things do:

• How close the software gets to the ceiling. The AMD boxes reach the iGPU through llama.cpp's Vulkan backend; the Intel Core Ultra goes through the SYCL / oneAPI path. Vulkan is the more mature of the two routes in llama.cpp today — the llama.cpp backend build docs list exactly what each backend supports. Same bus, different fraction of it actually used.
• Room to grow. Two SODIMM slots means 64 GB later while keeping dual-channel. One slot is a dead end, and the N100's single channel is precisely why it sits at less than half the bandwidth of everything else here.
• The escape hatch. The UM890 Pro exposes an OCuLink port; the SER8 and NUC 14 Pro give you USB4 / Thunderbolt 4. That is the difference between "this is as fast as it will ever be" and "I can attach a real GPU later" — see Thunderbolt vs OCuLink for local AI for how those two links differ.

The NPU on the Core Ultra is not a factor either way. Neither Ollama nor llama.cpp dispatches transformer inference to NPUs today, so treat it as future-proofing rather than a feature you can use.

Which mini PC should you actually buy?

If that last row is you, the budget local AI machine guide and used GPU buying guide cover the tower route. And if you like the form factor but need the bandwidth, Strix Halo mini PCs are the one small-form-factor class that genuinely changes the arithmetic.

What are the buying traps?

1. A single RAM stick

The most expensive mistake on this page, and it is invisible in the product title. One 32 GB SODIMM and two 16 GB SODIMMs are the same capacity and a factor of two apart in bandwidth: ~45 GB/s against ~90 GB/s, which is the difference between roughly 9 and roughly 18 tok/s on an 8B model. Insist on 2 × 16 GB or 2 × 32 GB.

2. Buying 16 GB and "upgrading later"

On 16 GB the OS and Ollama leave you around 10 GB for models — enough for an 8B at Q4 with a short context and little else. On single-slot boxes there is no "later" at all. Buy 32 GB in two sticks up front; our RAM requirements guide works through it size by size, and models for 8 GB systems covers what fits if you are already stuck.

3. Assuming the faster CPU is the faster box

The Ryzen 9 8945HS is a higher bin of the same silicon, with the same Radeon 780M and the same memory controller as the 8845HS. Generation speed is bandwidth-bound, so the ceiling does not move. Pay that premium for the OCuLink port or the dual NIC, not for tokens.

4. Sustained load in a small chassis

Mini PCs use low-profile coolers and a configurable TDP the vendor sets in firmware. Over a long inference session the chip settles nearer its sustained power limit than its boost figure — so when you see any performance number quoted anywhere, ask whether it was taken cold or after ten minutes under load.

5. NVMe placement

The M.2 slot in a mini PC is usually unventilated and sits close to the CPU cooler. Model loading is the one operation that reads tens of gigabytes in a single burst, which is exactly the workload NVMe thermal throttling targets. Fit a low-profile heatsink if the chassis has clearance.

6. Pulling models over Wi-Fi

A 20-40 GB download over Wi-Fi is a bad evening. Plug into Ethernet for the initial pull, then go wireless afterwards if you must.

7. Leaving suspend enabled

Default desktop settings suspend the machine after 20 minutes idle, killing a background Ollama server. On any box acting as a server:

sudo systemctl mask sleep.target suspend.target hibernate.target hybrid-sleep.target

How do you set up Ollama on a mini PC?

On AMD iGPU boxes the path that gets you closest to the bandwidth ceiling is Vulkan offload. On Ubuntu 24.04:

# 1. Update kernel and install Mesa/Vulkan
sudo apt update && sudo apt -y full-upgrade
sudo apt install -y mesa-vulkan-drivers vulkan-tools

# 2. Confirm the iGPU is visible
vulkaninfo --summary | head -30

# 3. Install Ollama
curl -fsSL https://ollama.com/install.sh | sh

# 4. Tell Ollama to use Vulkan
sudo systemctl edit ollama.service
# add:
# [Service]
# Environment="OLLAMA_NUM_GPU=999"
# Environment="OLLAMA_VULKAN=1"

sudo systemctl daemon-reload
sudo systemctl restart ollama

# 5. Pull and run
ollama pull llama3.1:8b-instruct-q4_K_M
ollama run llama3.1:8b-instruct-q4_K_M --verbose "Hello"

That --verbose flag is the important one: Ollama prints its own eval rate at the end of the response. That figure is measured on your hardware, by the runner itself, and it is the only token-rate number worth trusting. Compare it against the ceiling for your memory configuration in the table above. Landing somewhere under the ceiling is expected; landing at a small fraction of it almost always means Ollama fell back to CPU-only, so re-check that vulkaninfo sees the iGPU.

For the Intel NUC 14 Pro (Arc iGPU), the equivalent is the oneAPI / level-zero path:

sudo apt install -y intel-opencl-icd intel-level-zero-gpu
clinfo | grep "Device Name"   # confirm Arc shows up
# Set in service file:
# Environment="ONEAPI_DEVICE_SELECTOR=level_zero:gpu"

Our local AI benchmarking guide covers how to turn that eval rate into a repeatable number of your own, and the complete Ollama guide picks up from install through day-to-day use.

Frequently asked questions

Can a mini PC really run Llama 3.1 8B at usable speeds?

On paper, yes. Dual-channel DDR5-5600 gives about 90 GB/s and an 8B model at Q4_K_M occupies about 4.9 GB, so the arithmetic ceiling is roughly 18 tok/s. Real output falls below the ceiling, but the ceiling is comfortably above reading speed, which is the bar that matters for a chat assistant. 13B is tight, and 70B is not realistic on this class of machine.

AMD or Intel mini PC for Ollama?

Their ceilings are identical, because they run the same dual-channel DDR5-5600. The difference is how much of that ceiling the software reaches: AMD's Radeon 780M goes through llama.cpp's Vulkan backend, Intel's Arc through SYCL / oneAPI, and Vulkan is the better-trodden path today. Neither choice raises the ceiling — only the memory configuration does.

Do I need a discrete GPU if I have a mini PC?

Compare the bandwidth. An RTX 3060 12 GB is a published 360 GB/s against roughly 90 GB/s for a dual-channel mini PC — about a fourfold difference in headroom, plus dedicated VRAM the OS is not competing for. For 8B-class models on a single user, the mini PC is enough. For 13B and up, or concurrent users, it is not.

How much RAM do I need in a mini PC for Ollama?

32 GB, and the "two sticks" part matters more than the total. Two 16 GB modules give you the full 128-bit bus; one 32 GB module gives you the same capacity at half the bandwidth. 64 GB is worth it only if you want to keep two models resident at once.

Can I run multiple users from one mini PC?

One comfortably. Concurrent streams divide the same ~90 GB/s between them, so a second user does not get a second ceiling — both halve. For genuine multi-user serving you want a discrete GPU and a batching server; our Ollama multi-user guide covers the queueing side.

Does the Intel or AMD NPU help with Ollama?

Not currently. Ollama and llama.cpp do not dispatch transformer inference to NPUs, so the NPU sits idle during LLM work. NPUs remain useful for vision and audio pipelines — treat it as future-proofing, not a feature you can use today.

Can I add an external GPU to a mini PC?

The Minisforum UM890 Pro has OCuLink, which carries PCIe 4.0 x4 — around 8 GB/s. The SER8 and NUC 14 Pro offer USB4 / Thunderbolt 4, which tunnel PCIe rather than expose it directly. Both are narrower than a desktop PCIe 4.0 x16 slot at about 32 GB/s, and that gap shows up mainly when loading weights across the link rather than during generation once they are resident in the GPU's own VRAM.

How do I work out the ceiling for a model this page does not list?

Multiply the parameter count in billions by 0.6 to get its size at Q4_K_M, then divide your memory bandwidth by that. A 14B model is 14 × 0.6 ≈ 8.4 GB, so on a 90 GB/s mini PC the ceiling is 90 ÷ 8.4 ≈ 11 tok/s. Two numbers, one division, no benchmark required.

Conclusion

The honest version of this comparison is shorter than most: four of these five boxes share a token ceiling, so the decision is not about which CPU is fastest. It is about getting two RAM sticks instead of one, choosing the backend with the more mature software path, and deciding whether you want an escape hatch to a real GPU later.

Start with the Beelink SER8 unless you have a specific reason to deviate — OCuLink pushes you to the UM890 Pro, vendor support to the NUC 14 Pro, price to the GMKtec. Then pair it with our Ollama setup guide and check your own eval rate against the ceiling in the table above.