Can the Apple M4 Pro run Llama 3.1 8B?

Yes — the 4-bit build is a 4.9 GB download and fits in 24–64 GB unified memory. Expect roughly ~48 tok/s.

What is the biggest LLM the Apple M4 Pro can run?

Qwen3-Next 80B-A3B is the largest model in our catalog that fits (48.5 GB at 4-bit). Expect about ~128 tok/s.

How fast is the Apple M4 Pro for local LLMs?

Token generation is memory-bandwidth bound. At roughly 273 GB/s, the Apple M4 Pro generates about ~48 tok/s on an 8B-class model at 4-bit — speed scales inversely with model size.

Is Apple unified memory good for local LLMs?

Yes — the GPU can use nearly the whole memory pool, so a 32 GB Mac fits models that would need a very expensive dedicated GPU. Bandwidth, not capacity, is usually the speed limit.

← All modelsDEVICE CHECK

What LLMs can the Apple M4 Pro run?

The Apple M4 Pro has 24–64 GB unified memory and roughly 273 GB/s of memory bandwidth. Below is every model in our catalog that fits, with estimated generation speed. Biggest pick: Qwen3-Next 80B-A3B at ~128 tok/s.

Specifications

Memory24–64 GB unified memory

Bandwidth~273 GB/s

Memory typeUnified memory

Released2024-10

Models on the Apple M4 Pro

67 / 73 models

Model	Download (Q4)	24 GB	48 GB	64 GB	~Speed
Qwen3-Next 80B-A3BAlibaba	48.5 GB	Won't fit	Won't fit	Runs	~128 tok/s
Qwen3 Coder Next 80B-A3BAlibaba	48.5 GB	Won't fit	Won't fit	Runs	~128 tok/s
Llama 3.1 70BMeta	42.8 GB	Won't fit	Won't fit	Runs	~5 tok/s
Llama 3.3 70BMeta	42.8 GB	Won't fit	Won't fit	Runs	~5 tok/s
DeepSeek R1 70BDeepSeek	42.8 GB	Won't fit	Won't fit	Runs	~5 tok/s
Qwen 3.5 35B-A3BAlibaba	21.2 GB	Won't fit	Runs	Runs	~128 tok/s
Qwen 3.6 35B-A3BAlibaba	21.2 GB	Won't fit	Runs	Runs	~128 tok/s
Command R 35BCohere	21.2 GB	Won't fit	Runs	Runs	~11 tok/s
Qwen3-VL 32BAlibaba	20.0 GB	Won't fit	Runs	Runs	~12 tok/s
EXAONE 4.5 33BLG AI Research	20.0 GB	Won't fit	Runs	Runs	~12 tok/s
Qwen 3 32BAlibaba	19.9 GB	Won't fit	Runs	Runs	~12 tok/s
Qwen 2.5 Coder 32BAlibaba	19.9 GB	Won't fit	Runs	Runs	~12 tok/s
QwQ 32BAlibaba	19.9 GB	Won't fit	Runs	Runs	~12 tok/s
DeepSeek R1 32BDeepSeek	19.9 GB	Won't fit	Runs	Runs	~12 tok/s
Granite 4.0 H SmallIBM	19.4 GB	Won't fit	Runs	Runs	~43 tok/s
Nemotron 3 Nano 30B-A3BNVIDIA	19.2 GB	Won't fit	Runs	Runs	~106 tok/s
Gemma 4 31BGoogle	18.6 GB	Won't fit	Runs	Runs	~12 tok/s
Qwen 3 30B-A3BAlibaba	18.5 GB	Won't fit	Runs	Runs	~116 tok/s
Qwen3-VL 30B-A3BAlibaba	18.2 GB	Won't fit	Runs	Runs	~128 tok/s
Gemma 3 27BGoogle	16.6 GB	Runs	Runs	Runs	~14 tok/s
Qwen 3.5 27BAlibaba	16.4 GB	Runs	Runs	Runs	~14 tok/s
Qwen 3.6 27BAlibaba	16.4 GB	Runs	Runs	Runs	~14 tok/s
Gemma 4 26B A4BGoogle	15.3 GB	Runs	Runs	Runs	~101 tok/s
Mistral Small 3.1 24BMistral AI	14.6 GB	Runs	Runs	Runs	~16 tok/s
Devstral 24BMistral AI	14.6 GB	Runs	Runs	Runs	~16 tok/s
Magistral Small 1.2Mistral AI	14.6 GB	Runs	Runs	Runs	~16 tok/s
Devstral Small 2 24BMistral AI	14.6 GB	Runs	Runs	Runs	~16 tok/s
Codestral 22BMistral AI	13.5 GB	Runs	Runs	Runs	~17 tok/s
GPT-OSS 20BOpenAI	12.7 GB	Runs	Runs	Runs	~106 tok/s
Phi-4 Reasoning Vision 15BMicrosoft	9.1 GB	Runs	Runs	Runs	~26 tok/s
Qwen 3 14BAlibaba	9.0 GB	Runs	Runs	Runs	~26 tok/s
DeepSeek R1 14BDeepSeek	9.0 GB	Runs	Runs	Runs	~26 tok/s
Phi-4 14BMicrosoft	8.9 GB	Runs	Runs	Runs	~26 tok/s
Ministral 3 14BMistral AI	8.5 GB	Runs	Runs	Runs	~27 tok/s
OLMo 2 13BAi2	8.3 GB	Runs	Runs	Runs	~28 tok/s
Gemma 3 12BGoogle	7.4 GB	Runs	Runs	Runs	~31 tok/s
Mistral Nemo 12BMistral AI	7.4 GB	Runs	Runs	Runs	~31 tok/s
Gemma 4 12BGoogle	7.3 GB	Runs	Runs	Runs	~32 tok/s
Mellum 2 12B-A2.5BJetBrains	7.3 GB	Runs	Runs	Runs	~153 tok/s
Qwen 3.5 9BAlibaba	5.5 GB	Runs	Runs	Runs	~43 tok/s
GLM-4.6V-FlashZ.ai	5.5 GB	Runs	Runs	Runs	~43 tok/s
Qwen 2.5 VL 7BAlibaba	5.0 GB	Runs	Runs	Runs	~46 tok/s
Qwen 3 8BAlibaba	5.0 GB	Runs	Runs	Runs	~47 tok/s
Granite 3.3 8BIBM	5.0 GB	Runs	Runs	Runs	~47 tok/s
Llama 3.1 8BMeta	4.9 GB	Runs	Runs	Runs	~48 tok/s
DeepSeek R1 8BDeepSeek	4.9 GB	Runs	Runs	Runs	~48 tok/s
Gemma 4 E4BGoogle	4.9 GB	Runs	Runs	Runs	~85 tok/s
Qwen3-VL 8BAlibaba	4.9 GB	Runs	Runs	Runs	~48 tok/s
Ministral 3 8BMistral AI	4.9 GB	Runs	Runs	Runs	~48 tok/s
Gemma 3n E4BGoogle	4.7 GB	Runs	Runs	Runs	~96 tok/s
Qwen 2.5 Coder 7BAlibaba	4.6 GB	Runs	Runs	Runs	~50 tok/s
DeepSeek R1 7BDeepSeek	4.6 GB	Runs	Runs	Runs	~50 tok/s
Mistral 7BMistral AI	4.4 GB	Runs	Runs	Runs	~53 tok/s
Gemma 4 E2BGoogle	3.1 GB	Runs	Runs	Runs	~166 tok/s
Gemma 3 4BGoogle	2.6 GB	Runs	Runs	Runs	~89 tok/s
Qwen 3 4BAlibaba	2.4 GB	Runs	Runs	Runs	~96 tok/s
Qwen 3.5 4BAlibaba	2.4 GB	Runs	Runs	Runs	~96 tok/s
Phi-4 Mini 3.8BMicrosoft	2.3 GB	Runs	Runs	Runs	~101 tok/s
Llama 3.2 3BMeta	1.9 GB	Runs	Runs	Runs	~120 tok/s
DeepSeek-OCRDeepSeek	1.8 GB	Runs	Runs	Runs	~672 tok/s
Ministral 3 3BMistral AI	1.8 GB	Runs	Runs	Runs	~128 tok/s
DeepSeek R1 1.5BDeepSeek	1.1 GB	Runs	Runs	Runs	~213 tok/s
Qwen 3 1.7BAlibaba	1.0 GB	Runs	Runs	Runs	~225 tok/s
SmolLM2 1.7BHugging Face	1.0 GB	Runs	Runs	Runs	~225 tok/s
Llama 3.2 1BMeta	0.7 GB	Runs	Runs	Runs	~319 tok/s
Gemma 3 1BGoogle	0.6 GB	Runs	Runs	Runs	~383 tok/s
Qwen 3 0.6BAlibaba	0.4 GB	Runs	Runs	Runs	~638 tok/s

Apple Silicon shares one memory pool between CPU and GPU, so nearly all unified memory is available to the model (minus OS overhead). · Data updated: 2026-06-11 · How we calculate these numbers →