Short answer: for Ollama, buy the Mac mini M6 with 24GB if you mostly run 8B to 14B models, and the 32GB version if you want the 20B to 30B class models like gpt-oss:20b, gemma3:27b or qwen3:30b. Skip the 16GB base model for AI. It runs 8B models fine, but it has less memory bandwidth (153GB/s vs 170GB/s on Apple's spec sheet) and the memory can never be upgraded. If you need 70B models like llama3.3:70b, the M6 tops out at 32GB, so the only Mac mini that makes sense is the M5 Pro with 64GB.
Apple's current Mac mini comes in two chips: the new M6 (16GB, 24GB or 32GB of unified memory) and the M5 Pro (24GB, 48GB or 64GB). We checked every number below against Apple's own Mac mini tech specs page and Ollama's model library on Oct 27, 2026. No prices here because they change. Click through for today's price.
The Mac mini configs that matter for local AI
| Config (per Apple specs) | Unified memory | Memory bandwidth | Chip |
|---|---|---|---|
| M6 base | 16GB | 153GB/s | 12-core CPU, 12-core GPU, dual 16-core Neural Engine |
| M6 upgraded | 24GB or 32GB | 170GB/s | Same M6 chip |
| M5 Pro | 24GB, 48GB or 64GB | 307GB/s | 15-core CPU, 16-core GPU (configurable to 18-core CPU, 20-core GPU) |
Two things in that table decide everything for Ollama: how much memory (which models fit) and memory bandwidth (how fast the words come out). On Apple silicon the GPU and CPU share the same memory, so there is no separate VRAM. That is great for loading bigger models than a 12GB graphics card can hold, but every GB has to be bought on day one. Apple's spec sheet lists the memory as configurable at purchase only.
What each config actually runs in Ollama
Model sizes below are the download sizes Ollama lists for each default 4-bit tag. The model also needs room for its context window, and macOS keeps part of the memory for itself and caps how much the GPU can use. In practice, plan on roughly two thirds to three quarters of the total memory being available to the model.
| Model (Ollama tag) | Download size | M6 16GB | M6 24GB | M6 32GB | M5 Pro 64GB |
|---|---|---|---|---|---|
| llama3.1:8b | 4.9GB | Yes | Yes | Yes | Yes |
| qwen3:8b | 5.2GB | Yes | Yes | Yes | Yes |
| gemma3:12b | 8.2GB | Tight | Yes | Yes | Yes |
| qwen3:14b | 9.3GB | Tight | Yes | Yes | Yes |
| gpt-oss:20b | 14GB | No | Tight | Yes | Yes |
| mistral-small3.2 (24B) | 15GB | No | Tight | Yes | Yes |
| gemma3:27b | 17GB | No | No | Yes | Yes |
| qwen3:30b / qwen3-coder (MoE) | 19GB | No | No | Yes | Yes |
| qwen3:32b | 20GB | No | No | Tight | Yes |
| llama3.3:70b | 43GB | No | No | No | Yes |
"Tight" means it loads, but with a short context window and little left for your browser. If you see slowdowns, run ollama ps in Terminal. On a Mac you want the PROCESSOR column to read 100% GPU.
16GB: why the base model is a trap for AI
The 16GB M6 runs 7B to 8B models like llama3.1:8b and qwen3:8b well, and that is enough for chat, email drafts and simple summaries. The problem is everything after that. Once macOS, Safari and the model are loaded, a 12B to 14B model is already squeezed, and the 20B to 30B models that make local AI feel close to the paid chatbots do not fit at all. Apple's spec sheet also lists the 16GB model at 153GB/s of bandwidth, while the 24GB and 32GB configs get 170GB/s, so the base model is a little slower on every model too.
Buy 16GB only if local AI is a side experiment and you mainly want a quiet everyday desktop. If you already own a Windows PC, a 32GB RAM upgrade or a 12GB graphics card may get you further for less.
24GB: the sweet spot for most people
24GB is where a Mac mini starts to feel like a real local AI box. It runs every 8B to 14B model with room for a long context window and your normal apps, and it can squeeze in gpt-oss:20b. You also get the faster 170GB/s memory. For a family or small office chat assistant built on 8B to 14B models, this is the config we would buy.
32GB: the best M6 for local AI
32GB is the most memory the M6 offers, and it is the config to get if you want the stronger mid-size models: gemma3:27b, mistral-small3.2, qwen3:30b and qwen3-coder all fit with room to spare. The qwen3:30b family is a mixture of experts model, so it uses only a small part of its weights for each word and runs well on the M6's bandwidth. Since memory is fixed at purchase, 32GB is also the safest bet if you plan to keep the Mac mini for years while models keep growing.
M5 Pro 64GB: the only Mac mini for 70B models
A 4-bit llama3.3:70b is a 43GB download, so even 48GB is too tight once macOS and the context window take their share. The M5 Pro with 64GB is the only Mac mini config that runs 70B class models comfortably, and its 307GB/s bandwidth (per Apple's spec sheet) makes big models noticeably more usable than they would be on the M6. If 70B is your goal, also compare the GMKtec EVO-X2 with 128GB in our mini PC guide, which runs Windows.
Storage: do not forget the SSD
The base M6 ships with a 256GB SSD. Models add up fast: qwen3:30b alone is 19GB and llama3.3:70b is 43GB. If you plan to keep three or four models around, pick at least 512GB, or plan on an external SSD. The M6 has three Thunderbolt 4 ports on the back (up to 40Gb/s per Apple), which is plenty for a fast external drive to hold your model library.
How to set up Ollama on a Mac mini
- Download Ollama for macOS from ollama.com/download/mac and drag it to Applications.
- Open Terminal and run a model that fits your memory, for example
ollama run qwen3:14bon 24GB orollama run qwen3:30bon 32GB. - Run
ollama psto confirm it is running fully on the GPU. - Want a ChatGPT style window in your browser? Open WebUI works on a Mac too. The steps in our Ollama + Open WebUI guide are written for Windows, but the Open WebUI part is the same idea.
Mac mini vs a Windows PC for Ollama
Pick the Mac mini if you want a small, nearly silent box that runs 14B to 30B models out of the box with no driver hassles. Apple lists it at 5 by 5 inches and 2 inches tall. Pick a Windows PC if you already own a desktop tower, because a graphics card with 12GB to 16GB of VRAM is faster on models that fit and can be upgraded later. Not sure it is worth buying hardware at all? Run the numbers in our local AI vs cloud AI cost calculator.
Who should skip the Mac mini
- You rely on Windows-only software.
- You want to upgrade memory later. Every Mac mini config is fixed at purchase.
- You want the most speed per dollar on 8B to 14B models and already have a desktop tower. A GPU upgrade usually wins there.
FAQ
Is 16GB enough for Ollama on a Mac mini M6?
For 7B to 8B models like llama3.1:8b and qwen3:8b, yes. For 14B and bigger models it is tight or will not fit, and the memory cannot be upgraded later, so 24GB is the safer minimum for local AI.
Should I get 24GB or 32GB on the Mac mini M6?
Get 24GB if you mostly run 8B to 14B models. Get 32GB if you want 20B to 30B models such as gpt-oss:20b, gemma3:27b or qwen3:30b, or if you plan to keep the Mac mini for several years.
Can a Mac mini run Llama 3.3 70B?
Only the M5 Pro Mac mini with 64GB runs it comfortably. The 4-bit llama3.3:70b download is 43GB, and the M6 tops out at 32GB of unified memory.
Does Ollama use the GPU on a Mac mini?
Yes. Ollama uses the Apple silicon GPU automatically through the shared unified memory. Run ollama ps in Terminal and look for 100% GPU in the PROCESSOR column.
Can I upgrade Mac mini memory later?
No. Apple lists memory as a configure-to-order option at purchase only, so pick the amount you will need for the life of the machine.