Run Local LLMs on a Mac: The 2026 Guide (Hardware, Speeds, Tools)

The short answer: any Apple Silicon Mac with 16 GB of unified memory runs 7B–14B models well, a 64 GB M4 Pro runs 70B-class models at usable speeds, and a 256 GB Mac Studio M3 Ultra runs 200B-class models that no single consumer GPU can hold. Generation speed is set almost entirely by memory bandwidth, so the chip tier matters more than the year. Below: a hardware-to-model table, the RAM math, the three tools that matter (Ollama, LM Studio, MLX), a five-minute setup, and the honest case for renting a Mac instead of buying one.

Which Mac runs which model

The rule that decides everything: a model has to fit in unified memory at the quantization you choose, with a few gigabytes left for the context window and the OS. Once it fits, tokens per second scale with memory bandwidth. Our published benchmarks measured the base M4 (120 GB/s); the other rows are bandwidth-scaled estimates until we publish measured columns for them.

Per-model hardware requirements, quant tables, and the cheapest chip that fits each one live in our model guides — for example Llama 3.1 8B, Qwen 2.5 32B, and Llama 3.3 70B.

The RAM math in one line

Weights in GB ≈ parameters (billions) × bits per weight ÷ 8, plus about 15% runtime overhead, plus context. At Q4_K_M (≈4.85 bits) an 8B model is ~4.9 GB of weights and ~6 GB in memory; a 32B model is ~20 GB; a 70B model is ~42 GB; a 123B model is ~75 GB. Q8_0 roughly doubles those numbers, FP16 roughly quadruples them. Add 1–4 GB for a 32K context, more for 128K. That is why 16 GB tops out at 14B, 64 GB is the 70B threshold, and 128 GB is where 100B+ dense models become practical.

Why speed is about bandwidth, not cores

Generating one token means reading every active weight once. A 4.9 GB model on a 120 GB/s bus can therefore never exceed ~24 tok/s, and we measured 21.2 — about 87% of the theoretical ceiling. The same math predicts the M4 Pro (273 GB/s) at ~2× and the M4 Max (546 GB/s) at ~4× for the same model, which is what the community consistently reports. Two consequences: a bigger GPU core count on the same bandwidth barely helps generation, and mixture-of-experts models (which only read their active experts per token) run far faster than their parameter count suggests. Prompt processing is the exception — it is compute-bound, and we measured it varying 2× between model families at equal size (Qwen 2.5 7B 1,130 tok/s vs Llama 3.1 8B 587 tok/s).

Ollama vs LM Studio vs MLX vs llama.cpp

• Ollama — the default for anything headless or scripted: one command to pull a model, a local REST API on port 11434, and an OpenAI-compatible endpoint. It runs llama.cpp underneath. Hosting Ollama on a cloud Mac covers the always-on setup.
• LM Studio — the best GUI: model browser, chat window, and a local server that speaks the OpenAI API. Same engine class as Ollama, so same speeds; pick it if you want to click rather than type. Requirements and measured speeds in our LM Studio on a Mac guide.
• MLX— Apple's own array framework. On some models it is the fastest option on Apple Silicon and it is the natural choice for fine-tuning on a Mac; it is a Python library, not an app. See MLX on Apple Silicon.
• llama.cpp — the raw engine when you want every flag, the newest quant formats, or a C/C++ embed. llama.cpp on Apple Silicon has the Metal build and tuning notes.

Five-minute setup with Ollama

brew install ollama
ollama serve &                      # local API on :11434
ollama pull llama3.1:8b             # ~4.9 GB, fits any 16 GB Mac
ollama run llama3.1:8b "Explain unified memory in two sentences."

# OpenAI-compatible endpoint for your existing code:
curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"llama3.1:8b","messages":[{"role":"user","content":"Hi"}]}'

Point any OpenAI SDK at http://localhost:11434/v1 with a dummy API key and it works. To try a bigger model, check the RAM math above first — pulling a 70B build onto a 16 GB machine will download 40 GB and then fail to load.

Common mistakes

• Buying GPU cores instead of memory. A 24 GB Mac with more GPU cores runs the same 8B model no faster than a 16 GB one on the same chip; the next tier of bandwidth is what changes speed.
• Ignoring the context window. A model that “fits” with 1 GB to spare will swap and crawl the moment you paste a long document. Leave 2–4 GB free, more for RAG.
• Running production on a laptop. Thermal throttling, sleep, and a residential uplink turn a 21 tok/s machine into an unreliable one. Anything that needs to be up 24/7 belongs on a desktop-class Mac with a real network connection.

When a cloud Mac makes more sense than buying

If you need a local-class model available around the clock, reachable by your team or your app, the honest comparison is not laptop vs cloud but buying a Mac mini vs renting one. Owning wins on raw hardware cost over three years; renting wins the moment you count hosting, a static IP, power, remote-access setup, and being your own on-call. A dedicated M4 Mac mini from $79/mo comes with Ollama pre-installed and the OpenAI-compatible endpoint already exposed; the 64 GB M4 Pro tier is where 70B models live. If you would rather test before committing, deploy a model from the catalog — it takes about five minutes.