Overview
“Best” on this page means the largest Qwen2.5 Instruct model in the cited set whose 4-bit total from methodology is below the installed unified memory. It is not a quality ranking. No benchmark is used. Access date: 2026-10-12.
The 4-bit totals at context 4096, plus the assumed 2 GB overhead, are: 3B 3.696 GB, 7B 6.040 GB, 14B 10.155 GB, 32B 19.324 GB, 72B 39.692 GB. Parameter counts are 3.09B, 7.61B, 14.7B, 32.5B, and 72.7B on the model cards linked from how-much-ram-for-local-llm.
What is the best local LLM for a 16GB Mac?
On a 16 GB Mac the largest model in this set under the 4-bit total is Qwen2.5-14B-Instruct, at 10.155 GB. Qwen2.5-7B-Instruct totals 6.040 GB and leaves more of the 16 GB unused. Qwen2.5-32B-Instruct totals 19.324 GB, which is above 16 GB.
The 7B and 14B cards use the Apache-2.0 license. The 3B card, which also fits at 3.696 GB, uses the qwen-research license. Task quality is unknown. A longer context raises the 14B total: at 32768 tokens the 4-bit total is 15.792 GB, which is under 16 GB with a thin margin, and the macOS reserve is unknown.
Current Apple pages do not list an 8 GB configuration (see can-macbook-air-run-llm and which-mac-for-local-llm). LM Studio’s system requirements say you may still be able to use LM Studio on 8 GB Macs with smaller models and modest context sizes. That sentence is theirs. This page does not treat 8 GB as a current Apple SKU.
What can a 32GB or 36GB Mac run?
A 32 GB or 36 GB Mac is above the 32B 4-bit total of 19.324 GB and below the 72B 4-bit total of 39.692 GB.
Qwen2.5-14B-Instruct at 8-bit totals 17.505 GB (weights 14.7 + KV 0.805 + overhead 2), which is under 24 GB as well as under 32 GB. Qwen2.5-32B-Instruct at 8-bit totals 35.574 GB. That is under 36 GB by a fraction of a gigabyte and above 32 GB. Because the macOS reserve is unknown, this page does not call the 8-bit 32B model a fit on 36 GB.
At context 32768 the 32B 4-bit total is 26.840 GB (weights 16.25 + KV 8.590 + overhead 2), still under 32 GB. The 72B 4-bit total at 4096 is already 39.692 GB, so it is not under 32 or 36 GB.
What can a 64GB Mac run?
A 64 GB Mac is above the 72B 4-bit total of 39.692 GB. The 72B 8-bit weight size alone is 72.7 GB, which is above 64 GB, so the 8-bit 72B total is not under 64 GB.
The llama.cpp quantize README lists Llama 3.1 70B Q4_K_M at 43.1 GB. KV cache for that model is unknown here because the card was gated. 43.1 GB is under 64 GB and over 32 GB. Headroom after the file size, the KV cache, and macOS is unknown.
No mixture-of-experts card is cited on this page, so active-parameter versus total-parameter memory is unknown here. The quantization note on the sister site is llm quantization explained. Discrete GPU memory tiers are 24GB vs 16GB vs 12GB VRAM.
What changes at 128GB and above?
128 GB of installed memory is above the 72B 8-bit total of 76.042 GB (weights 72.7 + KV 1.342 + overhead 2). The MacBook Pro spec page lists 128 GB unified memory on the M5 Max with the 40-core GPU. The Mac Studio spec page lists 128 GB on the M5 Max with the 40-core GPU, and 256 GB or 512 GB on the M5 Ultra. Which of those configs you can order, and at what price, is not copied here.
This page does not name a 100B-class model. No such card is cited, so a 100B weight size is unknown. The formula on methodology can be applied when a card states a parameter count.
How do I pick between two models that both fit?
Pick the largest cited model whose total stays under your installed memory at the context you will use, then read the license on the model card. If two models both clear that bar, this page does not break the tie with a score. Open what-fits-my-ram and set the memory. Format choice is on mlx-vs-gguf. Install steps are on run-llm-on-mac.
Questions
Does a lower total mean a faster model?
Unknown. This page does not convert memory or bandwidth into tokens per second.
Why is 8 GB in the title?
People search for it. The Apple spec pages fetched on 2026-10-12 do not list 8 GB unified memory. The only vendor line quoted here is LM Studio’s note that 8 GB Macs may still run smaller models.