Which LLM on an iMac M4 (16 / 24 / 32 GB) ?

#What an iMac M4 can do for local AI

The iMac M4 runs 8- to 14-billion-parameter models well and accepts larger models at reduced speed. Three figures summarize the essentials. The M4 chip delivers 120 GB/s of memory bandwidth according to the Apple specifications, implying a theoretical ceiling of about 31 tokens per second on a 7B in Q4_0. The public llama.cpp repository benchmark on an M4 with a 10-core GPU reports 24.11 tokens per second during generation, or 77% of that ceiling. Finally, unified memory ranges from 16 to 32 GB and is chosen at purchase: it alone determines the possible model size. What this iMac lacks is bandwidth, not capacity.

Chip

Apple M4; the Apple specification does not offer a Pro or Max variant for the iMac.

Bandwidth

120 GB/s, regardless of the configuration.

Unified memory

16 GB standard; 24 GB optional on all versions, 32 GB on the 4-port model.

Screen

24-inch 4.5K (4480 × 2520), which leaves little reason to use a larger GPU: memory is the limit.

Pricing

Not indicated here: Apple modifies them; see the site's hardware page.

#Real-world configurations: two models, not three

The Apple page distinguishes two families. The 2-port model has an 8-core CPU and an 8-core GPU, with 16 GB of memory expandable to 24 GB. The 4-port model has a 10-core CPU and a 10-core GPU, with 16 GB expandable to 24 or 32 GB. The 120 GB/s bandwidth is identical. A common error, present in the older version of this page, is to assume that an 8 GB configuration still exists: the base memory is 16 GB. For an LLM, the key point is therefore to choose the 4-port model if you want 32 GB.

For text generation, the difference between 8 and 10 GPU cores is probably small: speed depends on bandwidth, which is identical. No public measurements from the reference table exist for the 8-core variant, so this claim remains a hypothesis. If you're unsure, pay for memory rather than cores.

#16, 24, or 32 GB: the memory budget

On a Mac, memory is shared between the system and the GPU, and macOS gives the GPU only a fraction of it. The iogpu.wired_limit_mb setting defaults to 0, meaning the kernel derives the limit from the installed memory. On a 32 GB Mac measured by ModelPiper, Metal made 24.96 GiB available to the GPU, or 78% of memory. 75% is commonly cited. For an iMac, therefore, expect roughly 11 to 12 GB for a 16 GB model, 17 to 18 GB for 24 GB, and 24 to 25 GB for 32 GB: these are estimates, so verify them on your machine with the sysctl iogpu.wired_limit_mb command and the Metal tool described by ModelPiper.

To go beyond these limits, you can increase the share allocated to the GPU. The guide to optimizing Macs with Apple Silicon details this setting and its risks; we do not repeat it here.

• Get the most out of a Mac Apple Silicon: macOS settings
• Memory calculator for a given model and context

#Which models for which memory capacity

The Q4 weights below come from the QuelLLM catalog. They assume the model fits in the usable memory shown in the previous table, including the context. The “smoothness” column applies the 120 GB/s ceiling to each model: these are calculated estimates, not measurements.

The key is not to confuse “fits” with “is comfortable.” A dense 24B model on an M4 base remains readable but slow, at a few words per second, whereas a mixture-of-experts (MoE) model such as gpt-oss 20B or Gemma 4 26B-A4B reads only a few billion parameters at each token, so it outpaces a comparably sized dense model. For a 24 or 32 GB iMac, an MoE is often the best choice.

A common trade-off on 24 GB is a 12B in Q8 (about 13 GB) or a 24B in Q4 (about 14 GB). Both use almost the same amount of memory, so they generate at roughly the same speed, around 7 tokens per second based on the bandwidth ceiling. The latter offers greater reasoning capacity, while the former stays closer to the original model. With equal bandwidth, model size in gigabytes determines throughput, not the number of parameters.

#Speed: bandwidth is the deciding factor

The only public reference measurement is the llama.cpp repository table on Apple Silicon, which tests a 7B LLaMA in Q4_0. For an M4 with a 10-core GPU and 120 GB/s, it reports 221,29 tokens per second for prompt processing and 24,11 for generation. An M4 Pro at 273 GB/s with a 16-core GPU reports 49,64 for generation, a little more than twice as much. This ratio is close to that of the memory bandwidths (2,3 times): a sign that generation depends on memory, not on the number of GPU cores.

These measurements date back to an early version of llama.cpp; recent versions and MLX may do better. Treat them as an order of magnitude. The previous version of this page reported 28 to 31 tokens per second for a 9B in Q5: that figure would exceed the 120 GB/s ceiling for a model of this size, so we removed it.

#Three realistic use cases based on memory

A memory budget is easier to understand when applied to a use case. In each case, add the model weights, context cache, and other open programs, then compare the total with the usable memory in the previous table. The site's calculator does this addition for you; just keep the logic in mind.

Regarding heat and noise, we found no public measurements of an M4 iMac under sustained LLM load, so we make no claims about its fan. The only reliable reference is actual usage: if generation runs for hours, monitor frequency and temperature in Activity Monitor and keep the context reasonable.

• Replacing ChatGPT with a local AI: what’s realistic

#Getting started

1. 01
   Check the memory

   Check your iMac's memory in the Apple menu, then About This Mac. It determines the list of possible models.
2. 02
   Install Ollama or LM Studio

   Ollama accelerates computations on Apple GPUs via the Metal API. LM Studio provides a GUI well suited to large screens.
3. 03
   Load a suitable model

   Choose a model from the table based on your memory, in Q4_K_M, and launch it with ollama run.
4. 04
   Monitor GPU usage

   Use ollama ps to verify that the model is loaded on the GPU, then use Activity Monitor to track memory pressure.
5. 05
   Limit context

   With 16 GB, keep the context to a few thousand tokens to prevent the system from swapping.

ollama ps
sysctl iogpu.wired_limit_mb
sysctl hw.memsize

• Ollama for beginners: installation and fundamentals
• LM Studio: complete tutorial

#iMac M4 or Mac mini M4: the real criterion

The Mac mini M4 uses the same chip and the same bandwidth; with the same amount of memory, the iMac is therefore not slower. What changes is the integrated, non-replaceable display and the lack of a Pro variant: the iMac tops out at 32 GB and 120 GB/s, while the Mac mini M4 Pro reaches 273 GB/s according to the same reference discussion. If your goal is a 24B or larger model at a reasonable speed, the Mac mini M4 Pro's bandwidth doubles the throughput; if it is a desktop assistant with an 8 to 14B model, the iMac is equivalent.

• Mac mini M4 and M4 Pro: which LLM for each memory size
• Mac mini M4 Pro hardware specifications
• iMac technical specifications, Apple
• llama.cpp performance on Apple Silicon, llama.cpp repository
• Ollama documentation: supported GPUs, including Metal
• ModelPiper: the GPU memory limit on Mac

#Frequently asked questions