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. 2049 of 2118 indexed models fit at 32K context with q4_0 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 1760vision language 185image 2audio asr 39audio tts 21video 16embedding 26

What fits at 32K context

largest quantization that fits, per model · 2049 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
GLM-4.5-Air-REAP-82B-A12BMoEQ4_081.9B45.78 GiB1.62 GiB47.98 GiB0.02 GiB22±37%
dolphin-2.6-mixtral-8x7bMoEQ8_046.7B46.22 GiB1.13 GiB47.93 GiB0.07 GiB13±37%
Nous-Hermes-2-Mixtral-8x7B-DPOMoEQ8_046.7B46.22 GiB1.13 GiB47.93 GiB0.07 GiB13±37%
Mixtral-8x7B-Instruct-v0.1MoEQ8_046.7B46.22 GiB1.13 GiB47.93 GiB0.07 GiB13±37%
xLAM-8x7b-rMoEQ8_046.7B46.22 GiB1.13 GiB47.93 GiB0.07 GiB13±37%
Open_Gpt4_8x7B_v0.1MoEQ8_046.7B46.22 GiB1.13 GiB47.93 GiB0.07 GiB13±37%
dolphin-2.5-mixtral-8x7bMoEQ8_046.7B46.22 GiB1.13 GiB47.93 GiB0.07 GiB13±37%
dolphin-2.7-mixtral-8x7bMoEQ8_046.7B46.22 GiB1.13 GiB47.93 GiB0.07 GiB13±37%
Mixtral-8x7B-v0.1MoEQ8_046.7B46.22 GiB1.13 GiB47.93 GiB0.07 GiB13±37%
Mixtral-8x7B-MoE-RP-StoryMoEQ8_046.7B46.22 GiB1.13 GiB47.93 GiB0.07 GiB13±37%
Noromaid-v0.4-Mixtral-Instruct-8x7b-ZlossMoEQ8_046.7B46.22 GiB1.13 GiB47.93 GiB0.07 GiB13±37%
Open_Gpt4_8x7B_v0.2MoEQ8_046.7B46.22 GiB1.13 GiB47.93 GiB0.07 GiB13±37%
Qwen3-Coder-REAP-25B-A3BMoEBF1624.9B46.34 GiB0.84 GiB47.73 GiB0.27 GiB24±37%
Hunyuan-A13B-InstructMoEQ4_K_L80.4B46.05 GiB1.13 GiB47.73 GiB0.27 GiB7±8.3%
CodeLlama-70b-Instruct-hfI1-Q5_K_S69.0B44.20 GiB2.81 GiB47.69 GiB0.31 GiB7±8.3%
CodeLlama-70b-Python-hfI1-Q5_K_S69.0B44.20 GiB2.81 GiB47.69 GiB0.31 GiB7±8.3%
Nous-Hermes-Llama2-70bI1-Q5_K_S69.0B44.20 GiB2.81 GiB47.69 GiB0.31 GiB7±8.3%
Midnight-Miqu-70B-v1.5I1-Q5_K_S69.0B44.20 GiB2.81 GiB47.69 GiB0.31 GiB7±8.3%
KafkaLM-70B-German-V0.1Q5_069.0B44.20 GiB2.81 GiB47.69 GiB0.31 GiB7±8.3%
llama2_70b_chat_uncensoredQ5_069.0B44.20 GiB2.81 GiB47.69 GiB0.31 GiB7±8.3%
Xwin-LM-70b-V0.1Q5_069.0B44.20 GiB2.81 GiB47.69 GiB0.31 GiB7±8.3%
Llama-2-70b-chat-hfQ5_069.0B44.20 GiB2.81 GiB47.69 GiB0.31 GiB7±8.3%
Llama-3_3-Nemotron-Super-49B-v1_5Q3_K_L49.9B24.47 GiB22.50 GiB47.66 GiB0.34 GiB7±8.3%
Valkyrie-49B-v2.1I1-Q3_K_L49.9B24.47 GiB22.50 GiB47.66 GiB0.34 GiB7±8.3%
Llama-3_3-Nemotron-Super-49B-v1Q3_K_L49.9B24.47 GiB22.50 GiB47.66 GiB0.34 GiB7±8.3%
Rombo-LLM-V3.0-Qwen-72bI1-Q4_K_M72.7B44.16 GiB2.81 GiB47.65 GiB0.35 GiB7±8.3%
Qwen2.5-72B-Instruct-abliteratedI1-Q4_K_M72.7B44.16 GiB2.81 GiB47.65 GiB0.35 GiB7±8.3%
Qwen2.5-72B-Instruct-abliterated-v2I1-Q4_K_M72.7B44.16 GiB2.81 GiB47.65 GiB0.35 GiB7±8.3%
HuatuoGPT-o1-72BQ4_K_M72.7B44.16 GiB2.81 GiB47.65 GiB0.35 GiB7±8.3%
MiroThinker-v1.0-72BI1-Q4_K_M72.7B44.16 GiB2.81 GiB47.65 GiB0.35 GiB7±8.3%
EVA-Qwen2.5-72B-v0.2Q4_K_M72.7B44.16 GiB2.81 GiB47.65 GiB0.35 GiB7±8.3%
Qwen2.5-Math-72B-InstructQ4_K_M72.7B44.16 GiB2.81 GiB47.65 GiB0.35 GiB7±8.3%
Qwen2.5-72B-InstructQ4_K_M72.7B44.16 GiB2.81 GiB47.65 GiB0.35 GiB7±8.3%
Malaysian-Qwen2.5-72B-InstructI1-Q4_K_M72.7B44.16 GiB2.81 GiB47.65 GiB0.35 GiB7±8.3%
Qwen2.5-72BI1-Q4_K_M72.7B44.16 GiB2.81 GiB47.65 GiB0.35 GiB7±8.3%
magnum-v4-72bI1-Q4_K_M72.7B44.16 GiB2.81 GiB47.65 GiB0.35 GiB7±8.3%
KAT-Dev-72B-ExpQ4_K_M72.7B44.16 GiB2.81 GiB47.65 GiB0.35 GiB7±8.3%
Homer-v1.0-Qwen2.5-72BQ4_K_M72.7B44.16 GiB2.81 GiB47.65 GiB0.35 GiB7±8.3%
Chuluun-Qwen2.5-72B-v0.01Q4_K_M72.7B44.16 GiB2.81 GiB47.65 GiB0.35 GiB7±8.3%
Qwen2.5-VL-72B-InstructQ4_K_M73.4B44.16 GiB2.81 GiB47.65 GiB0.35 GiB7±8.3%
Tower-Plus-72B-ultra-uncensored-hereticI1-Q4_K_M72.7B44.16 GiB2.81 GiB47.65 GiB0.35 GiB7±8.3%
Chronos-Platinum-72BQ4_K_M72.7B44.16 GiB2.81 GiB47.65 GiB0.35 GiB7±8.3%
UI-TARS-72B-DPOQ4_K_M73.4B44.16 GiB2.81 GiB47.65 GiB0.35 GiB7±8.3%
Step-3.5-Flash-REAP-121B-A11BI1-IQ3_XXS121B43.40 GiB3.67 GiB47.64 GiB0.36 GiB7±8.3%
Tess-3-Mistral-Nemo-12BF3212.2B45.63 GiB1.41 GiB47.64 GiB0.36 GiB7±8.3%
Lumimaid-v0.2-12BF3212.2B45.63 GiB1.41 GiB47.64 GiB0.36 GiB7±8.3%
MN-Violet-Lotus-12BF3212.2B45.63 GiB1.41 GiB47.64 GiB0.36 GiB7±8.3%
Mistral-Nemo-Instruct-2407F3212.2B45.63 GiB1.41 GiB47.64 GiB0.36 GiB7±8.3%
MN-12B-Celeste-V1.9F3212.2B45.63 GiB1.41 GiB47.64 GiB0.36 GiB7±8.3%
magnum-v2.5-12b-ktoF3212.2B45.63 GiB1.41 GiB47.64 GiB0.36 GiB7±8.3%
magnum-v2-12bF3212.2B45.63 GiB1.41 GiB47.64 GiB0.36 GiB7±8.3%
HunyuanImage-2.1Q5_017.5B47.04 GiB0.00 GiB47.64 GiB0.36 GiB7±8.3%
Qwen3-72B-SynthesisQ4_K_M72.7B44.12 GiB2.81 GiB47.61 GiB0.39 GiB7±8.3%
Behemoth-X-123B-v2IQ3_XXS123B43.78 GiB3.09 GiB47.58 GiB0.42 GiB7±8.3%
Mistral-Large-Instruct-2411IQ3_XXS123B43.78 GiB3.09 GiB47.58 GiB0.42 GiB7±8.3%
Qwen3.5-122B-A10B-hereticMoEI1-IQ3_XS123B46.72 GiB0.21 GiB47.51 GiB0.49 GiB36±37%
CalmeRys-78B-Orpo-v0.1I1-Q4_K_S78.0B43.72 GiB3.02 GiB47.43 GiB0.57 GiB7±8.3%
calme-2.3-rys-78bQ4_K_S78.0B43.72 GiB3.02 GiB47.43 GiB0.57 GiB7±8.3%
Qwen3.5-88BMoEI1-Q4_K_S87.7B46.63 GiB0.21 GiB47.41 GiB0.59 GiB33±37%
Huihui-Qwen3-Coder-Next-abliteratedMoEQ4_179.7B46.65 GiB0.21 GiB47.40 GiB0.60 GiB39±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?
2049 of 2118 indexed open-weight models fit a Apple M2 Max at 32,768 context with q4_0 KV cache, the largest being GLM-4.5-Air-REAP-82B-A12B 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.