Which LLMs Fit on a 64 GB Mac?

This is the size where 70B models fit. Llama 3.3 70B needs 50.0 GB at Q4_K_M with 8K, and Qwen 2.5 72B, the largest dense model that fits, needs 51.3 GB. Apertus v1.5 70B ranks first at 51.3 GB, a tight fit.

The largest model that fits is Nemotron 3.5 Bolt 74B (MoE), at 50.5 GB. It runs about 9B parameters per token, so it writes text much faster than a dense 70B model on the same Mac.

The order follows the model finder's best pick, which weighs size against release date. Treat the top rows as a short list to try, not a ranking of quality.

At 32K, the dense 70B models drop out. Llama 3.3 70B grows from 50.0 GB to 58.8 GB, since its KV cache takes 10.8 GB at 32K. A q8_0 KV cache halves that cache and brings it back under 56 GB, at a small cost in quality. You can set it in the calculator.

Nemotron 3.5 Bolt 74B barely changes. It needs 50.8 GB at 32K, because most of its layers keep a small fixed state instead of a KV cache. For long chats on this Mac, it's the big model to try first.

At Q8_0, the 30B class fits, even at 32K. K2 Horizon 32B, the largest dense model at Q8_0, needs 43.0 GB at 8K and 50.1 GB at 32K. The 70B models don't fit at Q8_0: Llama 3.3 70B needs 84.6 GB there.

So the choice is a 70B model at Q4_K_M with a short context, or a 30B model at Q8_0 with room for long chats. Nanbeige4.3 33B is a good middle ground, at 46.3 GB with Q8_0 and 32K.

The next models up are well past 56 GB. Llama 4 Scout needs 75.0 GB at Q4_K_M and still 60.6 GB at Q3_K_M. The 120B mixture of experts models, like gpt-oss 120B at 80.9 GB, are a size class up.

For those, see the 128 GB unified memory list. On a 64 GB Mac, a smaller quant won't close a gap this wide.

macOS doesn't let the GPU use all of a Mac's memory by default, and the default share can be less than this page assumes. On macOS Sonoma and later, you can raise the limit with a terminal command. It lasts until the next restart.

This sets it to the 56 GB this page uses:

sudo sysctl iogpu.wired_limit_mb=57344

Keep an eye on memory pressure in Activity Monitor. If the Mac starts swapping to disk, lower the limit or pick a smaller model.