Best AI mini PC for local LLMs
Last updated
The best AI mini PC for local LLMs depends on model size. Up to about 32B, we pick Olares One with a 24 GB RTX 5090 Mobile. For 70B to 120B, a 128 GB Ryzen AI Max+ 395 box or DGX Spark holds more.
Olares One → AI mini PC → CUDA · FLUX · Wan · Ollama → Image, video and chat on your desk
Which is the best AI mini PC for local LLMs at each model size?
Pick the cheapest box whose GPU memory holds your largest model. As an approximate guide, each billion parameters takes 0.5-0.6 GB at 4-bit, plus a few GB for context. Models up to about 32B fit in 24 GB of VRAM, and 70B and 120B models need 128 GB of shared memory.
| Model (4-bit) | Approx. memory | Mini PCs that fit |
|---|---|---|
| 8B to 14B | ~5 to ~9 GB | Every machine on this site |
| 32B | ~19-20 GB | Olares One, Mac mini (M5 Pro, 64 GB), all 128 GB boxes |
| 70B | ~40 GB | Ryzen AI Max+ 395 boxes, DGX Spark, ASUS GX10; Mac mini (64 GB) with little room |
| 120B | ~65 GB | Ryzen AI Max+ 395 boxes, DGX Spark, ASUS GX10 |
On shared-memory boxes the GPU uses only part of the total, so check the allocatable share for your model.
What decides speed once the model fits?
Memory bandwidth decides speed once a model fits, and mini PC makers publish it unevenly. NVIDIA lists 273 GB/s for DGX Spark and Apple lists 307 GB/s for the M5 Pro. GMKtec, Beelink and Minisforum give no figure for their Ryzen boxes, and Olares gives none for Olares One's VRAM.
So compare measured tokens/s for the model you plan to run. In Olares' own test, Olares One ran Qwen3-30B-A3B at 157 tokens/s for one request. That model is a mixture-of-experts design that uses only part of its weights per token, so a dense 32B model runs slower on any machine.
What happens when a model is bigger than the GPU?
On Olares One, the part of a model that does not fit in 24 GB of VRAM moves to the 96 GB of system RAM, and the CPU runs it. According to Olares' FAQ, gpt-oss-120b runs at about 36 tokens/s that way. That works for one person, but a 128 GB box holds such models entirely.
The 128 GB guide compares those boxes.
When should you skip our pick?
Skip Olares One if your main model is a dense 70B or your budget is below its 4,999 USD list price. In those cases a 128 GB Ryzen AI Max+ 395 box gives more memory for less money, and DGX Spark gives 128 GB with CUDA at 6,950 USD on NVIDIA's store (12 Nov 2026).
Keep it if the same box will also run image or video models or serve several people. It ships with Olares OS and installs Ollama and Open WebUI from the Olares Market. Compare the NVIDIA boxes in the NVIDIA mini PC guide and the software side in the Ollama guide.
Questions & answers
Can a mini PC run a 70B model?
Yes, if the GPU can reach enough memory. A dense 70B model needs about 40 GB at 4-bit by the approximate rule, so 128 GB Ryzen AI Max+ 395 boxes and DGX Spark hold it. Olares One's 24 GB of VRAM does not.
What is the fastest AI mini PC for a 30B model?
No cross-maker test is in our data, so we name none. In Olares' own test, Olares One ran Qwen3-30B-A3B at 157 tokens/s for one request and about 81 tokens/s per request with 8 concurrent requests.
Sources for this entry
- Olares blog: local AI hardware benchmarking (Oct 2025) (Company-published)
- Olares One FAQ (Olares docs) (Company-published)
- NVIDIA DGX Spark: maker's page (Maker-published)
- Apple Mac mini (M5 Pro, 64GB): maker's page (Maker-published)
- GMKtec EVO-X2 (Ryzen AI Max+ 395, 128GB): maker's page (Maker-published)
- Minisforum MS-S1 Max: maker's page (Maker-published)
- Beelink GTR9 Pro (Ryzen AI Max+ 395, 128GB): maker's page (Maker-published)
Continue this reading path
How the related websites connect
Cite this entry
Best AI Mini PC. “Best AI mini PC for local LLMs.” Content reviewed 13 November 2026.
https://bestaiminipc.com/best-ai-mini-pc-for-local-llm/