How much memory does a Mac need for local AI

A Mac can give its GPU only part of its memory, so the number on the box is not the number a model gets. By the rule macOS uses, a 16GB Mac offers the GPU about 10.7GB, a 32GB Mac about 21.3GB, and a 64GB Mac 48GB. That, plus a safety margin of 1 to 2GB, decides which open models run at full speed. For an 8B model, 16GB is enough. For a 32B model, you want 36GB. For a 70B model at 4-bit, 64GB is the least that works.

How much memory the GPU gets. A Mac's memory is shared between the CPU and the GPU. The macOS driver, as posted in llama.cpp discussion 2182, holds back 33.333 percent of the total when the Mac has 32GB or less, and 25 percent when it has more. So a 16GB Mac offers 10.7GB, a 24GB Mac 16.0GB, a 32GB Mac 21.3GB, a 36GB Mac 27GB, and a 128GB Mac 96GB. The same thread notes the limit can be raised with the sudo sysctl iogpu.wired_limit_mb command, at the cost of leaving less for macOS.

How much room to leave. The model's file is not all it needs. bartowski, who publishes most of the GGUF files people download, says on his model cards to pick a file 1 to 2GB smaller than the memory you have. In the table below, "fits" means at least 2GB spare, "tight" means 1 to 2GB, and "no" means less than 1GB or a file that is larger than the memory. The sums are ours. The file sizes are as listed on Hugging Face, read 19 October 2026.

Our AI model checker does this sum for any Mac, any graphics card or a memory size you type, for 20 open models.

16GB. The base Mac mini M6 at $899 and the MacBook Air have 16GB. It runs 7B and 8B models at any quantization, and a 14B model at 4-bit only just. This is the size for trying local AI, summarising notes and coding help from a small model.

24GB and 32GB. The Mac mini M6 can be configured to 24 or 32GB, and the Mac mini M5 Pro starts at 24GB for $1,699. At 24GB, models up to about 20B fit, including OpenAI's gpt-oss 20B at 12.1GB. At 32GB, Gemma 3 27B at 4-bit fits with room to spare, and Qwen3 32B is tight.

36GB. The Mac Studio M5 Max starts here, at $2,499. Both the 27B and 32B classes fit with room for a long context.

64GB. The M5 Pro tops out here. It is the least that runs a 70B model at 4-bit: Llama 3.3 70B at Q4_K_M is 42.5GB against 48GB of GPU memory. Q5_K_M, at 49.9GB, does not fit.

96GB and up. The Mac Studio M5 Ultra starts at 96GB for $5,499 and goes to 512GB. gpt-oss 120B at 63.4GB fits at 96GB. Only 128GB and up holds a 70B model at 8-bit (75.0GB).

The prices are Apple's list prices for the base configuration of each, as shown on our product pages. Memory upgrades cost extra, and we have not listed those prices.

Fitting is the first test. How fast it generates is the second, and that depends on memory bandwidth. In the community llama.cpp Apple Silicon scoreboard, a 7B model at 4-bit (Llama 2 7B Q4_0) generates 35.7 tokens per second on the M6, 67.6 on the M5 Pro, 119.9 on the M5 Max and 179.1 on the M5 Ultra. Apple lists their memory bandwidth as 153, 307, 460 and 1,200GB/s. The Max and Ultra figures are for versions with more GPU cores than the base Studio models, so a base Studio will be slower than those two.

A bigger model is slower than these figures: a 70B file is roughly ten times the size of a 7B one, and every token needs the whole file read. We have no published 70B speed for these Macs, so we will not give a number.

The short version is to buy for the largest model you will actually run. 16GB for 8B models, 32GB for the 20B to 27B class, 64GB for a 70B at 4-bit, and 96GB or more for the 100B-plus class. Past that, the buyer's guide says where each Mac stands in Apple's update cycle, and the Apple section has every cited benchmark.