Hardware · SoC / APU · Apple

Apple M3

Shares one pool of LPDDR5 memory between CPU and GPU at 100 GB/s. What fits depends on how much memory the system has.

What it runs

Largest models without offloading on the MacBook Air M3 16 GB, 8K context.

Variant
Qwen2.5 14B Instruct
14.8B · MLX 4-bit via MLX-LM
Tight fit
~7 tok/s estimate
Variant
Phi-4
14.7B · MLX 4-bit via MLX-LM
Tight fit
~7 tok/s estimate
Variant
Gemma 2 9B IT
9.24B · Q4_K_M via llama.cpp
Runs well
~11 tok/s estimate
Variant
Hermes 3 Llama 3.1 8B
8.03B · Q8_0 via llama.cpp
Tight fit
~7 tok/s estimate
Variant
Llama 3.1 8B Instruct
8.03B · Q8_0 via llama.cpp
Tight fit
~7 tok/s estimate
Variant
Qwen2.5 7B Instruct
7.62B · Q8_0 via llama.cpp
Runs well
~8 tok/s estimate
Variant
DeepSeek-R1-Distill-Qwen-7B
7.62B · Q4_K_M via llama.cpp
Runs well
~13 tok/s estimate
Variant
Mistral 7B Instruct v0.3
7.25B · Q8_0 via llama.cpp
Runs well
~8 tok/s estimate

System MacBook Air M3 16 GB
By estimate, runs models up to ~16B parameters at 4-bit in unified memory. What runs

Throughput on systems containing this device.

No performance measurements recorded yet.

Contributions are not open yet, so there is nothing here from members.

Contributions are not open yet, so there is nothing here from members.

No source records linked.