What fits in each amount of memory?
The model, its working memory for the conversation, and macOS all share one pool. On Apple silicon the GPU can use most of it, but not all: by default macOS caps the GPU’s share, so we size against that cap.
| Memory | Room for model + context (est.) | Fits for one user | Doesn’t fit well | Macs sold with it |
|---|---|---|---|---|
| 16GB | ~10–12GB | 8B models (qwen3:8b 5.2GB); gemma3:12b (8.2GB) with a short context | 14B models (phi4 9.1GB, qwen3:14b 9.3GB) leave almost no room | MacBook Air, iMac, Mac mini M6, MacBook Pro M6 |
| 24GB | ~16–18GB | 14B models (qwen3:14b 9.3GB, phi4 9.1GB); gpt-oss:20b (14GB) is tight | 24–27B models (mistral-small3.2 15GB, gemma3:27b 17GB) | MacBook Air, iMac, Mac mini, MacBook Pro |
| 32GB | ~21–24GB | gpt-oss:20b (14GB), mistral-small3.2 (15GB), gemma3:27b (17GB); qwen3:30b (19GB) with a short context | qwen3:32b (20GB) plus a useful context | MacBook Air, iMac, Mac mini M6, MacBook Pro M6 |
| 36GB | ~27GB | qwen3:32b (20GB), qwen3:30b (19GB) and gemma3:27b with room for longer documents | 70B models | MacBook Pro and Mac Studio with M5 Max (32-core GPU) |
| 48GB | ~36GB | 30B-class models for several people at once; qwen3:30b at 8-bit (33GB) for one | llama3.3:70b at 4-bit (43GB) | Mac mini M5 Pro, MacBook Pro M5 Pro or Max, Mac Studio M5 Max |
| 64GB | ~48GB | llama3.3:70b at 4-bit (43GB) for one user with a short context | gpt-oss:120b (65GB) | Mac mini M5 Pro, MacBook Pro M5 Pro or Max, Mac Studio M5 Max |
| 96GB | ~72GB | gpt-oss:120b (65GB); llama3.3:70b with longer context | qwen3:235b (142GB) | Mac Studio M5 Ultra |
| 128GB | ~96GB | gpt-oss:120b with room to spare; llama3.3:70b at 8-bit (75GB) | qwen3:235b (142GB) | MacBook Pro and Mac Studio with M5 Max (40-core GPU) |
| 256GB | ~192GB | qwen3:235b at 4-bit (142GB) | qwen3:235b at 8-bit (250GB) | Mac Studio M5 Ultra (80-core GPU) |
| 512GB | ~384GB | qwen3:235b at 8-bit (250GB) | Very little in Ollama’s library | Mac Studio M5 Ultra (80-core GPU) |
Memory options and bandwidth from Apple’s US spec pages for each Mac, read November 16, 2026. Model sizes are Ollama’s default 4-bit (Q4_K_M) downloads unless marked, read the same day. The GPU share is not published by Apple: about three-quarters of memory is what users report on 48GB and larger Macs, and about two-thirds to three-quarters on 32GB. Fit verdicts are our arithmetic for one user at about an 8,000-token context; more users or longer documents need more room.
Every 2026 Mac by memory and bandwidth
Read from Apple’s US spec pages on November 16, 2026. The MacBook Air uses the M5 chip; the iMac and the entry Mac mini and MacBook Pro use the M6.
| Mac | Chip | Memory you can order | Memory bandwidth |
|---|---|---|---|
| MacBook Air | M5 | 16, 24 or 32GB | 153GB/s |
| iMac | M6 | 16, 24 or 32GB | Up to 170GB/s |
| Mac mini | M6 | 16GB; 24 or 32GB | 153GB/s at 16GB; 170GB/s at 24 or 32GB |
| Mac mini | M5 Pro | 24, 48 or 64GB | 307GB/s |
| MacBook Pro | M6 | 16, 24 or 32GB | 153GB/s at 16GB; 170GB/s at 24 or 32GB |
| MacBook Pro | M5 Pro | 24 or 48GB (16-core GPU); 24, 48 or 64GB (20-core GPU) | 307GB/s |
| MacBook Pro | M5 Max | 36GB (32-core GPU); 48, 64 or 128GB (40-core GPU) | 460GB/s (32-core GPU); 614GB/s (40-core GPU) |
| Mac Studio | M5 Max | 36GB (32-core GPU); 48, 64 or 128GB (40-core GPU) | 460GB/s (32-core GPU); 614GB/s (40-core GPU) |
| Mac Studio | M5 Ultra | 96GB; 256 or 512GB with the 80-core GPU | 1.2TB/s |
Memory is chosen when you order these Macs and can’t be upgraded afterward, so buy for the largest model you expect to run.
How much does memory bandwidth matter?
For generating text, a lot. Each new word means reading the model’s weights from memory, so on the same model a Mac with twice the bandwidth generates roughly twice as fast. That is why a 307GB/s M5 Pro Mac mini and a 153GB/s Mac with the same 32GB can run the same model at very different speeds. Two other levers: a smaller model is faster than a bigger one, and a mixture-of-experts model such as qwen3:30b (about 3 billion parameters active per word) is much faster than a dense model of the same size.
Reading a long document before answering is a different job: it depends more on GPU cores. That is the case for the 20-core and 40-core GPU options.
What we don’t publish: tokens-per-second figures. Speeds depend on the model, the context length, the app and the number of people sharing the machine, and we only quote numbers we have measured on our own box.
Which Mac should you buy?
Trying local AI on your own
A 32GB MacBook Air, iMac or Mac mini runs 14B to 27B models for one person. 16GB works for 8B models and little else.
A shared server for a small office
A Mac mini with M5 Pro and 48GB or 64GB: 30B-class models for several people at once, at 307GB/s, in a small always-on box.
70B models or gpt-oss:120b
64GB is the floor for a 70B model at 4-bit, with little room to spare. For 120B-class models, a Mac Studio or MacBook Pro with M5 Max and 128GB, or a Mac Studio M5 Ultra.
The largest open models
qwen3:235b at 4-bit is 142GB, so only the Mac Studio M5 Ultra with 256GB or 512GB holds it.
Laptop or desktop for an office server?
A desktop. A server needs to stay awake, stay plugged in and sit somewhere safe, and a Mac mini or Mac Studio does that without a lid, a battery or someone taking it home. A MacBook Pro is the right buy when one person wants large models on the go. For an office, see the Mac mini AI server guide for setup, and local LLMs for business for which models suit which work.
The managed option: AdminOnsite installs a Mac mini sized to your office, preloaded and maintained, for a flat monthly price: Practice uses an M6 with 32GB, Office an M5 Pro with 48GB and Firm an M5 Pro with 64GB. See pricing.
What LLM can I run locally on a Mac?
It depends on memory. 16GB runs 8B models, 24GB runs 14B, 32GB runs 20–27B models such as gpt-oss:20b and gemma3:27b, 48–64GB runs 30B-class models for several people (and a 70B at 4-bit on 64GB), and 96GB or more runs gpt-oss:120b.
Can I run an LLM locally on a MacBook Pro?
Yes. The M6 MacBook Pro goes to 32GB, the M5 Pro to 64GB and the M5 Max to 128GB at 614GB/s, which is enough for gpt-oss:120b (65GB) or llama3.3:70b at 8-bit (75GB).
What is the best Mac for running local LLMs?
The one with enough memory for the model you want, then the most bandwidth you can afford. For most offices that is a Mac mini with M5 Pro and 48–64GB; for 120B-class models, an M5 Max with 128GB or a Mac Studio M5 Ultra.
Questions
Is 16GB enough for a local LLM?
For 8B models, yes. 14B models technically load but leave almost no room for the conversation, so 24GB or 32GB is the practical minimum for 14B.
Can a Mac run a 70B model?
At 4-bit, llama3.3:70b is 43GB, so it needs a 64GB Mac at minimum and fits best with 96GB or 128GB. It is a large dense model, so it generates much more slowly than 30B-class mixture-of-experts models on the same Mac.
Mac Studio or Mac mini for local AI?
The Mac mini tops out at 64GB and 307GB/s. The Mac Studio starts at 36GB with 460GB/s and goes to 512GB at 1.2TB/s. If your models fit in 64GB, the Mac mini is the smaller, cheaper box; past that, it has to be a Mac Studio.
Can I raise the GPU memory limit?
Users report a kernel setting that raises it, but Apple doesn’t document the default share or that setting for this purpose. We size against the default.
Ollama, LM Studio or MLX?
All three use the Apple GPU. Ollama is simplest as a shared server; LM Studio is a desktop app; MLX is Apple’s own framework. Model sizes in the table are Ollama’s.
Does AdminOnsite sell Macs?
No. AdminOnsite rents a preloaded, managed Mac mini to small offices for a flat monthly price, and it’s yours after 24 paid months.
Sources, checked November 16, 2026
- Apple: Mac mini technical specifications
- Apple: MacBook Pro technical specifications
- Apple: Mac Studio technical specifications
- Apple: MacBook Air technical specifications
- Apple: iMac technical specifications
- Apple Developer: Metal recommendedMaxWorkingSetSize
- LLM for Mac: Mac GPU memory limit (community reports, not Apple statements)
- Ollama: qwen3 tags and download sizes
- Ollama: gemma3 tags and download sizes
- Ollama: llama3.3 tags and download sizes
- Ollama: gpt-oss tags and download sizes
- Ollama: mistral-small3.2 tags and download sizes
- Ollama: phi4 tags and download sizes