Mac mini for LLM: Local Model Memory and Speed Calculator

Cloud Mac · By FideNubi Team · 2026-10-30 · Updated 2026-11-15

Short answer. A local LLM runs on a Mac mini when its model file, plus memory for the context, fits into the part of unified memory that macOS gives the GPU (about 75 % by default). In practice: a 16 GB Mac mini runs models up to about 12–14B parameters at 4-bit, such as Gemma 4 12B (about 11 GB); 24–32 GB runs the popular 20–35B models, such as gpt-oss 20B (about 16 GB, needs 24 GB) or Qwen3-Coder 30B (needs 32 GB); a 70B model needs the 64 GB Mac mini with an M4 Pro or M5 Pro, and even then it is tight. Pick a model below to see which cloud Mac offers can run it and what they cost.

Key facts

• Mac mini memory tops out at 32 GB with the M4 and M6 and at 64 GB with the M4 Pro and M5 Pro (Apple, Apple).
• Memory needed = model file + context cache + about 1 GB runtime; the file sizes come from Ollama, the context cache from each model's own configuration.
• Speed depends mostly on memory bandwidth: 307 GB/s (M5 Pro), 273 GB/s (M4 Pro), 170 GB/s (M6, 24/32 GB), 120 GB/s (M4), 100 GB/s (M2), per Apple.
• Mixture-of-experts models (gpt-oss, Qwen3-Coder 30B, Qwen 3.5 35B, Nemotron 3.5 Lightning) read only a few billion "active" parameters per token, so they run several times faster than dense models of the same size.
• Cheapest cloud Mac that runs gpt-oss 20B at our latest check: from €86 a month (Mac mini Vault, M2 24 GB).
• Catalog: 27 popular local language models in 58 sizes plus 19 models for images, video, speech, music and 3D, checked 2 November 2026; cloud Mac prices from our price tables.

Which cloud Mac mini can run your model?

Choose a task or a model, its size, the quantization and how much context you need. The calculator shows the memory needed, which Mac mini configurations can run it with an estimated speed, the cloud Mac offers from our price tables that fit (cheapest first, in euros per month), and the command to install it.

Not sure which model to pick? Best local LLM by task lists the models for coding, AI agents, documents, translation and RAG, and What LLM can a Mac mini run? starts from the Mac.

"Fits" means the model fits into the default GPU share of memory. "Tight" means it fits only after you raise the GPU memory limit (see below), leaving about 6 GB for macOS. Prices marked "~" are calculated from another billing period; "from" means the provider lists a starting price.

How much memory does a local LLM need on a Mac mini?

Every model and size in our catalog, at 4-bit (Q4) where Ollama offers it, with a 32K-token context: the memory it needs, the smallest Mac mini that runs it, and the cheapest cloud Mac with at least that much memory in our price tables.

A few things stand out. Small models (1–4B) fit in any Mac mini, but the context cache can be bigger than the model itself: Llama 3.2 3B needs more memory for a 32K conversation than for its weights. Mixture-of-experts models need memory for all their parameters but run fast. And the largest open models do not fit any Mac mini: gpt-oss 120B and Qwen 3.5 122B need more than 64 GB even at 4-bit, while 70B models (Llama 3.3, DeepSeek-R1 70B) fit the 64 GB Mac mini only tightly.

How do we calculate memory and speed?

We publish the formula so you can check it.

1. Model file. The real download size of the chosen build on Ollama (for example gpt-oss:20b, qwen3-coder:30b-a3b-q4_K_M). For Ollama's default tag we use the upper end of its size range, because on Apple silicon Ollama can serve a larger MLX build.
2. Context cache. For every token of context, a model keeps keys and values for each attention layer: 2 × key-value heads × head size × 2 bytes per layer. We read the layers and heads from the model's own config.json on Hugging Face. Layers with a sliding window only count up to their window, and linear-attention layers (Qwen 3.5 and 3.8) keep a small fixed state, so these models need far less memory for long contexts.
3. Runtime. About 1 GB for buffers.
4. What the Mac can give. macOS lets the GPU use only part of unified memory by default. Apple does not publish the share; community measurements put it near 75 %, which we use. It can be raised with sudo sysctl iogpu.wired_limit_mb=<megabytes> (resets at restart); we then keep 6 GB for macOS.
5. Speed. Each new token reads all active weights once, so the upper bound is memory bandwidth ÷ size of the active weights. We show 50–75 % of that bound as the estimate. For mixture-of-experts models we use the active size the developer states; where it is not stated, we show no speed.

Image, video, speech, music and 3D models work differently: generating an image or a video is limited by the GPU's compute rather than memory bandwidth, so the formula above does not apply. For these models the calculator shows the smallest Mac memory that the developer or users report as working, how the model runs on a Mac (supported by the developer, through a Mac app such as Draw Things, or only through a community port), and the speeds people measured on Macs, each with a link. Models grouped by these tasks are in Best local AI model by task.

The catalog is rebuilt from Ollama and Hugging Face whenever we update it, and checked monthly: new models are added, and models with a known macOS problem are left out (Laguna XS 2.1 is out for now: its Ollama page reports an issue on macOS).

Mac mini LLM performance: how fast each model runs

Speed scales with memory bandwidth, which Apple lists for each chip:

Apple does not list a bandwidth figure on the M1 Mac mini's spec page, so the calculator shows whether a model fits an M1 but no speed. Intel Mac minis run Ollama on the CPU only (Ollama) and are left out.

The difference between dense and mixture-of-experts models matters more than the chip. On the same M4 Pro, a dense 27B model at 4-bit reads about 17 GB per token, while Qwen3-Coder 30B reads only its 3.3B active parameters, a few gigabytes, so it answers several times faster.

How do you install a local LLM on a cloud Mac?

1. Connect to the cloud Mac. Providers give SSH access, screen sharing (VNC) or both; the calculator's offer list links to each provider's page.
2. Install Ollama. In the Terminal: curl -fsSL https://ollama.com/install.sh | sh, or download the installer from ollama.com. Ollama needs macOS Sonoma (14) or newer and uses both the CPU and the GPU on Apple silicon (Ollama).
3. Run the model. Copy the command from the calculator, for example ollama run gpt-oss:20b. The first run downloads the model file; after that it starts in seconds.
4. If the calculator says "tight". Raise the GPU memory limit before loading the model, for example sudo sysctl iogpu.wired_limit_mb=58000 on a 64 GB Mac mini. The setting resets when the Mac restarts.

Every model in the calculator links to its Ollama page and to the original release on Hugging Face, where you also find the licence: most are Apache 2.0 or MIT; Llama models use Meta's community licence and Gemma 3-based models Google's Gemma terms.

Rent or buy a Mac mini for local AI?

A cloud Mac lets you try a model on the exact chip and memory before buying, or run it around the clock without a machine at home. For agents such as OpenClaw, read Mac mini for OpenClaw and AI agents: rent, buy or VPS?; for 48 and 64 GB plans, Rent a Mac mini M4 Pro or M5 Pro in the cloud; and for every plan we track, the price tables for Europe, North America and Asia and Australia.

← All articles