Apple silicon / unified-memory buyer guide

Mac mini M4 for local LLMs: choose memory from the model, not the chip name.

Q: Can a 48GB Mac mini M4 Pro run a 32B local LLM?

The Qwen3 32B INT4 32K example reaches 27.93 GiB and fits within this page's 36 GiB planning budget for a 48GB configuration. That establishes capacity, not tokens-per-second performance.

Q: Is unified memory the same as GPU VRAM?

No. On Apple silicon, the CPU and GPU share unified memory. MLX arrays live in shared memory and can be used by either device without an explicit transfer, but macOS and other applications also consume the same pool. That is why this guide reserves more memory than the discrete-GPU pages.

Apple offers the 2024 Mac mini from 16GB to 64GB of unified memory across M4 and M4 Pro configurations. Because macOS, the CPU, and the GPU share that pool, the honest planning question is how much remains for weights, KV cache, and runtime allocations.

Test your model and memory →

Fast verdict

Sixteen starts the journey. Forty-eight changes the model class.

16GB installed
12 GiB planning budget

24GB installed
18 GiB planning budget

32GB installed
24 GiB planning budget

48GB / 64GB
36 / 48 GiB planned

Official configurations

M4 and M4 Pro do not offer the same memory ceiling.

Apple lists the M4 Mac mini with 16GB unified memory, configurable to 24GB or 32GB. The M4 Pro starts at 24GB and can be configured to 48GB or 64GB.

Configuration	Installed unified memory	75% planning budget	Capacity-oriented use
M4 base	16GB	12 GiB	Smaller quantized models and controlled context
M4 configured	24GB	18 GiB	14B INT4 and some short-context 20B plans
M4 configured	32GB	24 GiB	More margin, but not the shown 32B 32K plan
M4 Pro configured	48GB	36 GiB	The shown 32B INT4 32K plan clears the budget
M4 Pro configured	64GB	48 GiB	A dense 70B INT4 weight floor plus headroom clears on paper

The table is a memory-capacity map. It does not predict prompt-processing speed, decode speed, thermals, or the quality of a particular quantization.

Transparent memory checks

Match the configuration to a complete plan.

The linked calculator was designed around GPU memory. Here its single-device capacity field is used only as an editable proxy for installed unified memory, with a stricter 75% usable-memory setting. It does not model macOS memory pressure, swap, or MLX performance.

Qwen3 8B · INT4 · 32K · 16GB9.98 GiBFits the 12 GiB budget → Qwen3 14B · INT4 · 32K · 24GB14.27 GiBFits the 18 GiB budget → gpt-oss-20b · MXFP4 · 4K · 24GB14.29 GiBFits the 18 GiB budget → Qwen3 32B · INT4 · 32K · 32GB27.93 GiBMisses the 24 GiB budget → Qwen3 32B · INT4 · 32K · 48GB27.93 GiBFits the 36 GiB budget → 70B dense · INT4 floor + 20% · 64GB39.12 GiBCapacity floor only; add the real cache →

Unified memory is different

Do not copy a discrete-GPU VRAM table onto a Mac.

On a discrete NVIDIA card, model weights and cache compete inside a dedicated VRAM pool while system RAM remains separate. On Apple silicon, CPU and GPU workloads share unified memory with macOS and every other running application. Installed memory is therefore not an isolated accelerator budget.

MLX documents the benefit of this design: arrays live in shared memory and operations can run on the CPU or GPU without an explicit device-to-device transfer. That architectural advantage does not make all installed memory safely available to one model. Browser tabs, development tools, the inference server, and the operating system still need room.

This guide uses 75% as a conservative comparison convention. It leaves 4GB on a 16GB machine, 6GB on 24GB, 8GB on 32GB, 12GB on 48GB, and 16GB on 64GB outside the model plan. Your measured safe percentage may differ. Watch memory pressure and swap during the actual workload, then adjust the calculator rather than treating 75% as a platform rule.

Choose the memory first

Memory is not upgradeable later.
Select the installed pool around the largest model and context you genuinely expect to use.

Capacity is not throughput.
A plan that fits can still be too slow for interactive work; benchmark the exact runtime.

Swap is not free memory.
A model surviving with swap does not mean it meets your latency or system-stability target.

Primary sources and method

Separate Apple's facts from ResearchAudio's planning assumptions.

Apple supplies the available memory configurations and memory bandwidth figures. The MLX project documents the shared-memory execution model. ResearchAudio supplies the editable 75% reserve and scenario arithmetic. No tokens-per-second claim is made because speed depends on the exact chip, model, quantization, runtime, prompt length, and decode settings.

Official Apple Mac mini (2024) technical specifications → Official MLX project and unified-memory description → Start from your installed memory →

Is a 16GB Mac mini M4 enough for local LLMs?

It is a workable starting point for smaller quantized models. The Qwen3 8B INT4 32K example reaches 9.98 GiB and fits the conservative 12 GiB planning budget.

Should I choose 24GB or 32GB unified memory for local LLMs?

Choose 24GB for the shown 14B and short-context 20B plans. Choose 32GB for more operating margin, context, or concurrent apps. Neither clears the shown 32B 32K plan under the 75% reserve.

Can a 48GB Mac mini M4 Pro run a 32B local LLM?

The Qwen3 32B INT4 32K plan reaches 27.93 GiB and clears the 36 GiB planning budget. This proves arithmetic capacity, not speed.

Is unified memory the same as GPU VRAM?

No. Apple silicon shares unified memory across the CPU, GPU, operating system, and applications. This guide therefore uses a larger reserve than its discrete-GPU pages.