§01·spec · /gpus

Apple M3 Max

appleapple series48GB unified

21 open-weights AI models run on the Apple M3 Max — see which fit, how fast they go, and the VRAM each needs.

Apple Silicon uses unified memory shared between CPU and GPU. By default the GPU can address about 75% of it; the limit is raisable via Metal’s wired-memory setting.

§02·models that run on this GPU

21 total

LLM · 13

Model	Min VRAM	Works	Evidence
Devstral Small 2 (24B)	48GB	✓	recipe	check ↗
Gemma 4 12B	48GB	✓	recipe	check ↗
gpt-oss 20B	13GB	✓	recipe	check ↗
Laguna XS 2.1	24GB	✓	recipe	check ↗
Llama 3.1 8B	5GB	✓	recipe	check ↗
Llama 3.3 70B	40GB	✓	recipe	check ↗
Mistral Nemo 12B	48GB	✓	recipe	check ↗
Mistral Small 3.2 24B	48GB	✓	recipe	check ↗
North Mini Code 1.0	48GB	✓	recipe	check ↗
Ornith 1.0 35B	48GB	✓	recipe	check ↗
Phi-4	48GB	✓	recipe	check ↗
Qwen3 32B	19GB	✓	recipe	check ↗
Qwen3-Next 80B-A3B	48GB	✓	recipe	check ↗

Multimodal · 1

Model	Best speed	Min VRAM	Works	Evidence
Gemma 4 E4B-IT		5GB	✓	recipe	check ↗

Image · 3

Model	Min VRAM	Works	Evidence
Krea 2	24GB	✓	recipe	check ↗
Qwen-Image	23GB	✓	recipe	check ↗
Z-Image Turbo	17GB	✓	recipe	check ↗

Video · 1

Model	Best speed	Min VRAM	Works	Evidence
LTX-2.3		34GB	✓	recipe	check ↗

3D · 1

Model	Best speed	Min VRAM	Works	Evidence
TRELLIS.2-4B		17GB	✓	recipe	check ↗

TTS · 2

Model	Best speed	Min VRAM	Works	Evidence
Kokoro TTS		1GB	✓	recipe	check ↗
VoxCPM		2GB	✓	recipe	check ↗

§03·tested recipes

showing 6 of 21

llmadvanced48GB+recipe
Qwen3-Next 80B-A3B on Apple M3 Max: an 80B MoE Assistant in 48GB via a Sub-Q4 GGUF
llmintermediate48GB+recipe
Gemma 4 12B on Apple M3 Max: Local Private Assistant via llama.cpp / Ollama (48GB)
llmintermediate48GB+recipe
Phi-4 (14B) on Apple M3 Max: Full-Precision Local Assistant via llama.cpp / Ollama (48GB)
llmintermediate48GB+recipe
Mistral Nemo 12B on Apple M3 Max (48GB): Full-Precision Local Assistant via llama.cpp / Ollama (Metal)
llmintermediate48GB+recipe
Mistral Small 3.2 24B on M3 Max (48GB): Local Private Assistant via llama.cpp / Ollama on Apple Metal
llmintermediate48GB+recipe
Devstral Small 2 (24B) on Apple M3 Max: Local Agentic Coding via llama.cpp Metal + OpenHands (48GB Apple / Q8_0 near-lossless)

→ view all 21 recipes for Apple M3 Max

Apple M3 Max

Qwen3-Next 80B-A3B on Apple M3 Max: an 80B MoE Assistant in 48GB via a Sub-Q4 GGUF

Gemma 4 12B on Apple M3 Max: Local Private Assistant via llama.cpp / Ollama (48GB)

Phi-4 (14B) on Apple M3 Max: Full-Precision Local Assistant via llama.cpp / Ollama (48GB)

Mistral Nemo 12B on Apple M3 Max (48GB): Full-Precision Local Assistant via llama.cpp / Ollama (Metal)

Mistral Small 3.2 24B on M3 Max (48GB): Local Private Assistant via llama.cpp / Ollama on Apple Metal

Devstral Small 2 (24B) on Apple M3 Max: Local Agentic Coding via llama.cpp Metal + OpenHands (48GB Apple / Q8_0 near-lossless)