Apple · apple

Apple M2 Max

Apple M2 Max has 64 GB of unified memory at 410 GB/s — about 44.64 GiB usable after driver and compositor overhead. 2031 of 2118 indexed models fit at 32K context with f16 KV. Note only 48 GB of its 64 GB is allocatable to the GPU.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
64 GB
LPDDR5-6400
Bandwidth
410 GB/s
512-bit bus
Tensor FP16
—
dense
TDP
—
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1743vision language 184image 2audio tts 21video 16audio asr 39embedding 26

What fits at 32K context

largest quantization that fits, per model · 2031 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Hunyuan-A13B-InstructMoEQ4_080.4B43.44 GiB4.00 GiB47.98 GiB0.02 GiB7±8.3%
L3.3-Electra-R1-70bIQ4_NL70.6B37.30 GiB10.00 GiB47.98 GiB0.02 GiB7±8.3%
L3.3-70B-Magnum-v4-SEIQ4_NL70.6B37.30 GiB10.00 GiB47.98 GiB0.02 GiB7±8.3%
Hermes-4-70BIQ4_NL70.6B37.30 GiB10.00 GiB47.98 GiB0.02 GiB7±8.3%
Llama-3.3-70B-Instruct-abliteratedIQ4_NL70.6B37.30 GiB10.00 GiB47.98 GiB0.02 GiB7±8.3%
Llama-3.3-70B-InstructIQ4_NL70.6B37.30 GiB10.00 GiB47.98 GiB0.02 GiB7±8.3%
Llama-3.1-70BIQ4_NL70.6B37.30 GiB10.00 GiB47.98 GiB0.02 GiB7±8.3%
L3.3-70B-Euryale-v2.3IQ4_NL70.6B37.30 GiB10.00 GiB47.98 GiB0.02 GiB7±8.3%
Rombos-LLM-70b-Llama-3.3IQ4_NL70.6B37.30 GiB10.00 GiB47.98 GiB0.02 GiB7±8.3%
Anubis-70B-v1.2IQ4_NL70.6B37.30 GiB10.00 GiB47.98 GiB0.02 GiB7±8.3%
DeepSeek-R1-Distill-Llama-70BIQ4_NL70.6B37.30 GiB10.00 GiB47.98 GiB0.02 GiB7±8.3%
DeepSeek-R1-Distill-Llama-70B-abliteratedIQ4_NL70.6B37.30 GiB10.00 GiB47.98 GiB0.02 GiB7±8.3%
Legion-V2.1-LLaMa-70BIQ4_NL70.6B37.30 GiB10.00 GiB47.98 GiB0.02 GiB7±8.3%
Tess-R1-Limerick-Llama-3.1-70BIQ4_NL70.6B37.30 GiB10.00 GiB47.98 GiB0.02 GiB7±8.3%
SEMIKONG-70BIQ4_NL70.6B37.30 GiB10.00 GiB47.98 GiB0.02 GiB7±8.3%
Athene-70BIQ4_NL70.6B37.30 GiB10.00 GiB47.98 GiB0.02 GiB7±8.3%
Hermes-3-Llama-3.1-70BIQ4_NL70.6B37.30 GiB10.00 GiB47.98 GiB0.02 GiB7±8.3%
L3.3-70B-Magnum-DiamondIQ4_NL70.6B37.30 GiB10.00 GiB47.98 GiB0.02 GiB7±8.3%
Meta-Llama-3-70B-Instruct-abliterated-v3.5IQ4_NL70.6B37.30 GiB10.00 GiB47.98 GiB0.02 GiB7±8.3%
Meta-Llama-3-70B-InstructIQ4_NL70.6B37.30 GiB10.00 GiB47.98 GiB0.02 GiB7±8.3%
GLM-4.6VMoEQ2_K108B41.64 GiB5.75 GiB47.97 GiB0.03 GiB17±37%
Qwen3.5-88BMoEI1-Q4_K_S87.7B46.63 GiB0.75 GiB47.95 GiB0.05 GiB30±37%
Huihui-Qwen3-Coder-Next-abliteratedMoEQ4_179.7B46.65 GiB0.75 GiB47.94 GiB0.06 GiB36±37%
GPT-NeoX-20B-ErebusI1-Q5_K_M20.6B14.24 GiB33.00 GiB47.88 GiB0.12 GiB7±8.3%
Noromaid-20b-v0.1.1I1-IQ3_M20.0B8.53 GiB38.75 GiB47.87 GiB0.13 GiB7±8.3%
command-a-plus-05-2026-bf16MoEIQ1_S219B45.87 GiB1.42 GiB47.85 GiB0.15 GiB25±37%
Llama-4-Scout-17B-16E-Instruct-abliterated-v2MoEKV unresolvedI1-IQ3_XS109B41.25 GiB6.00 GiB47.83 GiB0.17 GiB17±37%
Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16MoEQ5_K_M35.1B46.64 GiB0.63 GiB47.82 GiB0.18 GiB33±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEIQ1_M124B44.51 GiB2.75 GiB47.81 GiB0.19 GiB24±37%
Mistral-Small-4-119B-2603MoEIQ3_XXS119B46.44 GiB0.70 GiB47.73 GiB0.27 GiB33±37%
Kimi-Dev-72BIQ4_XS72.7B37.02 GiB10.00 GiB47.70 GiB0.30 GiB7±8.3%
Qwen2.5-VL-72B-InstructIQ4_XS73.4B37.02 GiB10.00 GiB47.70 GiB0.30 GiB7±8.3%
GLM-4.5-AirMoEUD-IQ2_M110B41.34 GiB5.75 GiB47.67 GiB0.33 GiB17±37%
Rombo-LLM-V3.0-Qwen-72bI1-IQ4_XS72.7B36.98 GiB10.00 GiB47.66 GiB0.34 GiB7±8.3%
Qwen2.5-72B-Instruct-abliteratedI1-IQ4_XS72.7B36.98 GiB10.00 GiB47.66 GiB0.34 GiB7±8.3%
Qwen2.5-72B-Instruct-abliterated-v2I1-IQ4_XS72.7B36.98 GiB10.00 GiB47.66 GiB0.34 GiB7±8.3%
HuatuoGPT-o1-72BIQ4_XS72.7B36.98 GiB10.00 GiB47.66 GiB0.34 GiB7±8.3%
MiroThinker-v1.0-72BI1-IQ4_XS72.7B36.98 GiB10.00 GiB47.66 GiB0.34 GiB7±8.3%
EVA-Qwen2.5-72B-v0.2IQ4_XS72.7B36.98 GiB10.00 GiB47.66 GiB0.34 GiB7±8.3%
Qwen2.5-Math-72B-InstructIQ4_XS72.7B36.98 GiB10.00 GiB47.66 GiB0.34 GiB7±8.3%
Qwen2.5-72B-InstructIQ4_XS72.7B36.98 GiB10.00 GiB47.66 GiB0.34 GiB7±8.3%
Malaysian-Qwen2.5-72B-InstructI1-IQ4_XS72.7B36.98 GiB10.00 GiB47.66 GiB0.34 GiB7±8.3%
Qwen2.5-72BI1-IQ4_XS72.7B36.98 GiB10.00 GiB47.66 GiB0.34 GiB7±8.3%
magnum-v4-72bI1-IQ4_XS72.7B36.98 GiB10.00 GiB47.66 GiB0.34 GiB7±8.3%
KAT-Dev-72B-ExpIQ4_XS72.7B36.98 GiB10.00 GiB47.66 GiB0.34 GiB7±8.3%
Homer-v1.0-Qwen2.5-72BIQ4_XS72.7B36.98 GiB10.00 GiB47.66 GiB0.34 GiB7±8.3%
Tower-Plus-72B-ultra-uncensored-hereticI1-IQ4_XS72.7B36.98 GiB10.00 GiB47.66 GiB0.34 GiB7±8.3%
Chronos-Platinum-72BIQ4_XS72.7B36.98 GiB10.00 GiB47.66 GiB0.34 GiB7±8.3%
UI-TARS-72B-DPOIQ4_XS73.4B36.98 GiB10.00 GiB47.66 GiB0.34 GiB7±8.3%
HunyuanImage-2.1Q5_017.5B47.04 GiB0.00 GiB47.64 GiB0.36 GiB7±8.3%
Apertus-70B-Instruct-2509Q3_K_L70.6B36.87 GiB10.00 GiB47.60 GiB0.40 GiB7±8.3%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ2_K_L109B40.97 GiB6.00 GiB47.55 GiB0.45 GiB17±37%
Qwen3-Coder-Next-REAMMoEI1-Q6_K60.3B46.22 GiB0.75 GiB47.51 GiB0.49 GiB34±37%
Chuluun-Qwen2.5-72B-v0.01Q3_K_L72.7B36.79 GiB10.00 GiB47.47 GiB0.53 GiB7±8.3%
Qwen3-72B-SynthesisQ3_K_L72.7B36.79 GiB10.00 GiB47.47 GiB0.53 GiB7±8.3%
Behemoth-X-123B-v2IQ2_S123B35.75 GiB11.00 GiB47.45 GiB0.55 GiB7±8.3%
Delphi-25B-SimpleRL-MathI1-Q4_K_S25.0B13.35 GiB33.47 GiB47.44 GiB0.56 GiB7±8.3%
Huihui-GLM-4.7-Flash-abliterated-57BMoEI1-Q6_K57.3B42.64 GiB4.18 GiB47.42 GiB0.58 GiB19±37%
Nethena-20BQ3_K_S20.0B8.06 GiB38.75 GiB47.41 GiB0.59 GiB7±8.3%
Laguna-S-2.1MoEUD-IQ3_S118B45.10 GiB1.64 GiB47.32 GiB0.68 GiB28±37%
●From the file◐Predictedwhat these mean

Speed is modeled, not measured: decode is memory-bandwidth bound, so tokens per second is bytes read per token against achievable bandwidth. Mixture-of-experts models carry a wider band because only the routed experts are read each step, and few have been measured publicly.

Questions people ask

What AI models can a Apple M2 Max run?
2031 of 2118 indexed open-weight models fit a Apple M2 Max at 32,768 context with f16 KV cache, the largest being Hunyuan-A13B-Instruct at Q4_0. That covers text, vision-language, image, video and speech models.
How much usable memory does a Apple M2 Max actually have?
Its nameplate is 64 GB, but about 44.64 GiB is available to a model once driver and compositor overhead is accounted for, and only 48 GB of the pool can be allocated to the GPU at all.
Is a Apple M2 Max fast for local AI?
Its memory bandwidth is 410 GB/s, and that figure — not teraflops — is what governs token generation speed. Capacity decides what you can run; bandwidth decides how fast it runs.