Best Mini PCs for Local Inference: Running 7B-13B LLMs Without a Tower

Running a language model on your own hardware used to mean building a tower with a discrete GPU, a loud fan curve, and a power bill to match. For a large swath of local-AI work, that is now overkill. If your goal is a private chat assistant, a coding helper, or a summarizer running a quantized 7B-13B model, a small AMD mini PC with 32 GB of unified memory does the job on a desk corner, silently, for under a thousand dollars.

The catch is that these boxes live and die by one number: system RAM. There is no discrete GPU with its own VRAM here. The integrated Radeon 780M graphics and the CPU share the same 32 GB pool, so that pool is your entire model-and-context budget. Understand that constraint and a mini PC is a genuinely good local-inference machine. Ignore it and you will be disappointed the first time you try to load something a mini PC was never meant to run.

This guide covers three closely matched AMD mini PCs built around the Radeon 780M iGPU and 32 GB of DDR5. They are the sweet spot for quantized small-model inference via tools like Ollama and llama.cpp. Below we break down where each one earns its price and where the whole category tops out.

How we picked

• Workload fit for small quantized models. Every pick targets the same job: running 4-bit quantized 7B-13B LLMs on CPU plus the Radeon 780M iGPU. That is chat, coding assistance, and summarization, not model training and not 70B-class inference.
• Verified specs only. We list the CPU, iGPU, memory, and storage exactly as specified for each unit. We do not quote NPU throughput or synthetic benchmark scores, because we do not have verified figures for these units and will not invent them.
• 32 GB unified memory as the floor. All three ship with 32 GB of DDR5. That is the practical minimum for comfortably holding a 13B model plus a working context window, since the iGPU shares that same memory.
• Value at the sub-$1,000 tier. These are close on paper, so we weight price and what you get for it heavily.
• In-stock on Amazon. Each pick was live and buyable when we captured pricing. Prices are list prices captured in July 2026 and change frequently, so verify before buying.

The OneClickAI Score

Our proprietary editorial composite, disclosed in full so you can judge it yourself:

OneClickAI Score = Capability (40) + Value (30) + Real-World Fit (20) + Build & Support (10). Each sub-score is our editorial assessment on a 0-100 scale within its category, then weighted.

These sub-scores are our judgment based on the verified specs and the target workload. They are not lab measurements, and we do not pretend they are. The Score column is the exact weighted average of the four sub-scores.

Product	Capability	Value	Real-World Fit	Build & Support	Score
Beelink SER8	84	92	88	85	87.3
GEEKOM A8 Max	90	82	88	86	86.8
MINISFORUM UM890 Pro	88	85	86	84	86.3

The spread is deliberately tight because these machines are genuinely close. The Beelink edges ahead on pure value, the GEEKOM leads on raw capability with the strongest CPU, and the MINISFORUM sits between them. None of these gaps should override a good sale price on the day you buy.

Beelink SER8 — the value pick

Specs: AMD Ryzen 7 8745HS (8 cores / 16 threads, Zen 4, 4nm), Radeon 780M integrated graphics, 32 GB DDR5, 1 TB PCIe SSD. List price around $889.

The SER8 is the machine we would hand to most people getting started with local inference. It runs a tier below the other two on CPU, the Ryzen 7 8745HS instead of a Ryzen 9, but it pairs the same Radeon 780M iGPU with the same 32 GB of DDR5 and a full 1 TB PCIe SSD, and it does so for the lowest price in this group.

For the actual workload here, that CPU-tier difference matters less than the price gap suggests. A 4-bit quantized 7B or 13B model is bottlenecked far more by the shared 32 GB memory pool and memory bandwidth than by the last increment of CPU clock. You get the same memory ceiling as the pricier boxes, the same iGPU, and a terabyte of fast storage for model files, which fill up quickly when you keep several quantized models on hand.

Who it's for: Anyone standing up their first local assistant or coding helper who wants the most local-inference machine per dollar and does not need the top CPU bin.

Pros:

• Lowest price of the three at roughly $889
• Full 1 TB PCIe SSD for holding multiple model files
• Same 32 GB unified-memory ceiling and Radeon 780M iGPU as the more expensive picks

Cons:

• Ryzen 7 8745HS is a step below the Ryzen 9 units for CPU-bound work
• 32 GB is a hard ceiling shared with the iGPU, same as every box here

Check price on Amazon

GEEKOM A8 Max — the capability pick

Specs: AMD Ryzen 9 8945HS (8 cores / 16 threads), Radeon 780M integrated graphics, 32 GB DDR5, 1 TB SSD. List price around $999.

The A8 Max steps up to the Ryzen 9 8945HS, the strongest CPU in this roundup, while keeping the familiar Radeon 780M iGPU, 32 GB of DDR5, and a 1 TB SSD. It is the most expensive of the three at around $999, and the premium buys you the top CPU bin.

Where does that extra CPU headroom show up? In the parts of local inference that lean on the processor: prompt processing on longer inputs, running the model on CPU cores when you are not offloading layers to the iGPU, and general responsiveness when the box is also doing other work. It will not change the ceiling on model size, because that is still the 32 GB memory pool, but for the same small quantized models it has a bit more compute in reserve.

Who it's for: Buyers who want the strongest CPU in the category and are comfortable paying the most for it, especially if the mini PC will double as a general workstation between inference sessions.

Pros:

• Ryzen 9 8945HS, the top CPU tier here
• 1 TB SSD plus 32 GB DDR5, matching the storage and memory of the value pick
• Same quiet, low-power Radeon 780M platform

Cons:

• Highest price of the three at roughly $999
• No larger memory ceiling than the cheaper boxes, so the premium is CPU-only

Check price on Amazon

MINISFORUM UM890 Pro — the middle ground

Specs: AMD Ryzen 9 8945HS (8 cores / 16 threads), Radeon 780M integrated graphics, 32 GB DDR5. List price around $949.

The UM890 Pro splits the difference. It carries the same Ryzen 9 8945HS as the GEEKOM but lists at around $949, below the A8 Max. It keeps the Radeon 780M iGPU and 32 GB of DDR5, so its local-inference ceiling is identical to the other two.

If you want the Ryzen 9 CPU without paying the full A8 Max price, this is the box that gets you there. One note: our verified spec sheet for this unit lists the CPU, iGPU, and 32 GB of memory, but not a storage capacity. Confirm the SSD size in the current Amazon listing before you buy, since storage headroom matters when you are keeping several multi-gigabyte quantized models around.

Who it's for: Buyers who want the Ryzen 9 8945HS but not the top-of-range price, and who will verify the storage configuration before ordering.

Pros:

• Ryzen 9 8945HS at a lower price than the GEEKOM
• Same Radeon 780M iGPU and 32 GB unified-memory platform as the rest
• Sits neatly between the value and capability picks

Cons:

• Storage capacity is not confirmed in our verified specs; check the listing
• Same 32 GB ceiling; the Ryzen 9 does not lift the model-size limit

Check price on Amazon

Quick comparison

Product	Key spec	Price	Best for	Score
Beelink SER8	Ryzen 7 8745HS, 32 GB DDR5, 1 TB SSD	~$889	Best value entry point	87.3
GEEKOM A8 Max	Ryzen 9 8945HS, 32 GB DDR5, 1 TB SSD	~$999	Top CPU in the group	86.8
MINISFORUM UM890 Pro	Ryzen 9 8945HS, 32 GB DDR5	~$949	Ryzen 9 at a lower price	86.3

Prices are list prices captured in July 2026 and change frequently — check the current price on Amazon before buying.

Local inference buying guidance

What size model can a 32 GB mini PC run?

The honest answer is small, quantized models: 7B-13B parameters at 4-bit quantization, run through Ollama or llama.cpp. That range fits comfortably in 32 GB alongside a working context window. Remember that on these machines the 32 GB is unified memory shared between the CPU and the Radeon 780M iGPU, so it is not 32 GB of model space plus separate graphics memory. It is one pool for everything. That is why 32 GB is the floor we recommend rather than a comfortable margin: it is enough for a 13B model with room to work, not enough to jump to much larger models.

Mini PC or a GPU for local AI?

It depends entirely on the model size you are targeting. A mini PC like these is quieter, cheaper, and far lower on power draw than a GPU tower, and for 7B-13B assistants and coding helpers that trade-off is very much worth making. But a mini PC is much slower on large models and is not a substitute for a 24 GB discrete GPU when you want to run 70B-class work. If 70B inference is your goal, look at a dedicated GPU build instead. If you want a compact, always-on box for small models, the mini PC wins on noise, cost, and power. Our GPUs for local LLMs guide covers the other side of that fork.

Does the iGPU help?

The Radeon 780M is an integrated GPU, so it shares the same system memory rather than bringing its own VRAM. Tools like Ollama and llama.cpp can put part of the workload on the iGPU, which helps, but the iGPU is drawing from the same 32 GB pool as everything else. Treat the 780M as a useful assist for small quantized models, not as a replacement for the dedicated VRAM you would get on a discrete card. The ceiling is still that shared 32 GB of unified memory.

How much storage do I actually need?

Quantized models are multi-gigabyte files, and it is normal to keep several around to compare. Two of these picks ship with 1 TB SSDs, which gives you real room. The MINISFORUM's storage is not confirmed in our verified specs, so check its listing. If you plan to hoard models, a fast portable SSD for AI models is a cheap way to add capacity without opening the case.

Bottom line

For most people getting into local inference, the Beelink SER8 is the pick. It matches the memory ceiling, iGPU, and 1 TB storage of the pricier boxes for the lowest price, and the CPU-tier gap barely registers for quantized 7B-13B work that is bounded by memory, not clock speed. That is why it takes the top OneClickAI Score.

Step up to the GEEKOM A8 Max if you specifically want the Ryzen 9 8945HS and will use the machine as a general workstation between inference sessions. Consider the MINISFORUM UM890 Pro if you want that same Ryzen 9 for less than the GEEKOM, and confirm its storage before you order.

Whichever you choose, buy it for what it is: a quiet, efficient box for small quantized models. If your ambitions run to 70B models, that is a different machine. Plan the rest of the setup with our budget local-AI build hub and the full complete AI hardware stack guide.

Prices are list prices captured in July 2026 and change frequently — check the current price on Amazon before buying.