How much of a Mac's unified memory can the GPU use for LLMs?

Where the limit comes from

The limit is Metal's recommendedMaxWorkingSetSize, which Apple describes as “an approximation of how much memory, in bytes, this GPU device can allocate without affecting its runtime performance”. In its Metal Compute on MacBook Pro tech talk Apple gave two values: on an M1 Pro or M1 Max with 32 GB the GPU can access 21 GB, and on an M1 Max with 64 GB, 48 GB.

llama.cpp reads that property as the Metal device's working-set size (ggml-metal-device.m:1289-1290) and reports free memory as the limit minus what is already allocated (1501-1511). The same source notes that it is possible to allocate more than the limit, and llama.cpp warns when it does (2009-2021), so the limit is a recommendation that apps follow rather than a hard wall. MLX uses the same value as its system wired limit (mlx memory.cpp).

Apple does not publish a rule for every memory size. The site uses two thirds up to 32 GB and three quarters above, which matches Apple's two figures and the public llama.cpp logs in the table below. Every Mac page, the LLM VRAM calculator and the 2026 VRAM report use this same rule.

Raising the limit with iogpu.wired_limit_mb

sysctl iogpu.wired_limit_mb              # 0 means the macOS default
sudo sysctl iogpu.wired_limit_mb=24576   # 24 GB for the GPU on a 32 GB Mac
sudo sysctl iogpu.wired_limit_mb=0       # back to the default

• Units are MiB. On a 32 GB Mac, 24576 took llama.cpp's reported limit from 22,906.50 MB to 25,769.80 MB, which is 24 GiB (peddals.com, community test).
• It lasts until the next restart. The change applies at once; LM Studio shows the new value after a relaunch. A restart returns to the default, and the same post shows how to keep it with /etc/sysctl.conf, with a warning to do that carefully.
• macOS version. iogpu.wired_limit_mb is the name on macOS 14 Sonoma and later; on 13 Ventura it was debug.iogpu.wired_limit, and 0 means the default split (llama.cpp #4406, community). Besides community posts, the sources for it are the MLX project's API docs and the mlx-lm README, which tell users to set it to a value above the model's size and below the machine's memory.
• Risk. Memory the GPU wires is not available to macOS and your apps. Leave several GB, quit big apps and save work first. One user with a 24 GB M4 Pro running large MoE models at a limit of 22000 reported kernel panics (llama.cpp #19825).

For a single Mac, its page on this site shows the exact command for the raised limit and which models it adds, for example the M1 Max or M2 Max Mac (32 GB) or the M5 Max Mac (128 GB).

Reading the real limit in the llama.cpp log

llama.cpp and apps built on it print the limit when the Metal device starts (ggml-metal-device.m:1362-1366):

ggml_metal_device_init: recommendedMaxWorkingSetSize  = 22906.50 MB

• Current builds print bytes ÷ 1,000,000. Divide by 1,073.74 for GiB: 22,906.50 MB is 21.33 GiB, two thirds of 32 GB.
• The model loader then reports that limit minus what is already allocated as “MiB free” for the Metal device.
• When a buffer takes the total past the limit, llama.cpp prints current allocated size is greater than the recommended max working set size; at debug log level it also logs every Metal buffer as “allocated / limit” in MiB (2009-2021).
• Ollama writes the same line to its server log in ~/.ollama/logs/ (peddals.com).
• In MLX, device_info() reports it as max_recommended_working_set_size.

To check a model against your Mac before downloading it, use the LLM VRAM calculator, which counts the Mac at the default limit, and the LLM speed calculator for tokens per second from the Mac's memory bandwidth.