Can I run Code Llama 34B on a Mac?

Code Llama 34B needs about 20.5 GB of memory. It fits on 12 of the 15 machines we track. The cheapest current machine that runs it at a comfortable 10+ tokens/sec is the Mac Studio M5 Max (40-core GPU, 64GB) ($3,799).

How much memory does Code Llama 34B need?

About 20.5 GB: 19 GB for the 4-bit model plus room for context.

Can I run Code Llama 34B locally? Speed on every Mac & mini PC

Machine	Memory	Fits?	Tokens/sec	From
Mac Studio M5 Ultra (96GB)	96 GB	Yes	52	$5,499	Check price →
Mac Studio M5 Ultra (256GB)	256 GB	Yes	52	$9,499	Check price →
Mac Studio M5 Max (40-core GPU, 64GB)	64 GB	Yes	26	$3,799	Check price →
Mac mini M5 Pro (64GB)	64 GB	Yes	13		Check price →
Mac mini M4 Pro (64GB) older	64 GB	Yes	12		Check price →
NVIDIA DGX Spark (128GB)	128 GB	Yes	10		Check price →
Ryzen AI Max+ 395 mini PC (128GB)	128 GB	Yes	9.4		Check price →
Mac mini M5 Pro (24GB)	24 GB	Barely	7.8	$1,699	Check price →
Mac mini M6 (32GB)	32 GB	Yes	7.2		Check price →
Beelink SER7 (Ryzen 7 7840HS, 32GB)	32 GB	Yes	2.8		Check price →
Minisforum UM790 Pro (Ryzen 9 7940HS, 32GB)	32 GB	Yes	2.8		Check price →
GEEKOM A6 (Ryzen 7 6800H, 32GB)	32 GB	Yes	2.4		Check price →
Mac mini M6 (16GB)	16 GB	No	–	$899	Check price →
Mac mini M4 (16GB) older	16 GB	No	–		Check price →
Mac mini M2 (16GB) older	16 GB	No	–		Check price →

Around 10 tokens/sec reads comfortably; under 5 feels slow. How we estimate.

Other codellama sizes

Code Llama 7B Code Llama 13B Code Llama 70B

Embed this table

Free for blogs, READMEs and model cards. Updates automatically.

<iframe src="https://vsmacs.com/embed/codellama-34b" width="100%" height="420" style="border:0" loading="lazy" title="Code Llama 34B speed by machine"></iframe>