Best local LLM for a 32GB Mac

Which models fit on a 32GB Mac?

Anything up to about 24 GB, including room for the conversation. macOS keeps part of the 32GB for itself, so our finder counts three quarters of it as usable for model weights (our estimate). That means dense models up to roughly 35 billion parameters at 4-bit, or bigger mixture-of-experts models with a small active part.

Why mixture-of-experts models win on a base 32GB Mac

Speed on a Mac is mostly about memory bandwidth, because every new token reads the model's active weights once. A base M5 moves 153 GB/s. Qwen3.8 27B is dense, so each token reads all 27.8B parameters, about 20 GB at Q5. That works out to roughly 5 tokens per second on an M5: it fits, but it crawls.

Qwen3.6 35B A3B only reads its 3B active parameters per token, so the same chip runs it at about 51 tokens per second (our estimates). That's the whole trick for base M4, M5 and M6 Macs. On an M1 or M2 Max, with 400 GB/s, dense 27B to 31B models get to a usable speed and their quality starts to pay off.

What won't fit, and is 32GB enough?

These are the closest misses on a 32GB Mac:

• Recursive Flint 40B A2.5B needs about 25 GB even at Q4.
• Llama 3.3 70B Instruct needs about 44 GB even at Q4.
• gpt-oss-120b needs about 72 GB even at Q4.

Compared with 24GB, 32GB is a real step up. A 24GB Mac leaves about 16 GB for a model by our estimate, and the largest model on our list that fits is gpt-oss-20b. At 32GB the 30B-class models all come into play. The next jump is 48GB (about 36 GB usable), where Recursive Flint 40B A2.5B fits.

If you're buying new, note the gap in Apple's lineup. On the Mac mini, 32GB is only offered with the base M6 chip, and the M5 Pro model goes from 24GB straight to 48GB or 64GB. The MacBook Air with M5 also tops out at 32GB. If local models are the main reason you're buying, the 48GB M5 Pro is the better buy; if they're a side interest, a 32GB M5 or M6 runs every pick on this page.

How to run these models on a Mac

The two easiest options are free. With Ollama, install the app and pull a model from the Ollama library with one command. With LM Studio, search the model name and pick the quant from the table above. On Apple silicon, LM Studio's MLX builds are usually the fastest.

Two habits help on 32GB. Close memory-hungry apps before loading a 20 GB model, and keep the context window modest, since a long conversation needs memory on top of the model itself. If a pick feels slow, drop to the next model down in the table rather than a lower quant.

Not sure local is worth it versus paying per token? Our cheapest LLM API guide shows what the same work costs in the cloud, and best models for coding covers the hosted options.