Apple · apple

Apple M2 Max

Apple M2 Max has 96 GB of unified memory at 410 GB/s — about 66.96 GiB usable after driver and compositor overhead. 2064 of 2118 indexed models fit at 64K context with q8_0 KV. Note only 72 GB of its 96 GB is allocatable to the GPU.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
96 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 1774vision language 186image 2audio asr 39audio tts 21video 16embedding 26

What fits at 64K context

largest quantization that fits, per model · 2064 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
MiniMax-M2.7MoEIQ2_XS229B63.21 GiB8.23 GiB71.98 GiB0.02 GiB13±37%
Step-3.7-FlashIQ2_S201B57.93 GiB13.30 GiB71.80 GiB0.20 GiB5±8.3%
Qwen3-VL-235B-A22B-ThinkingMoEUD-IQ1_M236B64.90 GiB6.24 GiB71.73 GiB0.27 GiB13±37%
Qwen3-VL-235B-A22B-InstructMoEUD-IQ1_M236B64.83 GiB6.24 GiB71.65 GiB0.35 GiB13±37%
MiMo-V2.5MoEKV unresolvedIQ1_M311B67.01 GiB3.98 GiB71.59 GiB0.41 GiB18±37%
GLM-4.7-REAP-218B-A32BMoEIQ2_XS218B58.74 GiB12.22 GiB71.55 GiB0.45 GiB9±37%
GLM-4.5-Air-DerestrictedMoEQ4_1110B64.77 GiB6.11 GiB71.46 GiB0.54 GiB13±37%
GLM-4.5-AirMoEQ4_1110B64.77 GiB6.11 GiB71.46 GiB0.54 GiB13±37%
Mistral-Medium-3.5-128BQ3_K_M128B58.94 GiB11.69 GiB71.33 GiB0.67 GiB5±8.3%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ4_1109B64.35 GiB6.38 GiB71.30 GiB0.70 GiB13±37%
HuatuoGPT-o1-72BQ6_K72.7B59.93 GiB10.63 GiB71.23 GiB0.77 GiB5±8.3%
Rombo-LLM-V3.0-Qwen-72bQ6_K72.7B59.93 GiB10.63 GiB71.23 GiB0.77 GiB5±8.3%
Qwen2.5-72B-Instruct-abliteratedQ6_K72.7B59.93 GiB10.63 GiB71.23 GiB0.77 GiB5±8.3%
EVA-Qwen2.5-72B-v0.2Q6_K72.7B59.93 GiB10.63 GiB71.23 GiB0.77 GiB5±8.3%
MiroThinker-v1.0-72BQ6_K72.7B59.93 GiB10.63 GiB71.23 GiB0.77 GiB5±8.3%
Qwen2.5-Math-72B-InstructQ6_K72.7B59.93 GiB10.63 GiB71.23 GiB0.77 GiB5±8.3%
Qwen2.5-72B-InstructQ6_K72.7B59.93 GiB10.63 GiB71.23 GiB0.77 GiB5±8.3%
Qwen2.5-72BQ6_K72.7B59.93 GiB10.63 GiB71.23 GiB0.77 GiB5±8.3%
Kimi-Dev-72BQ6_K72.7B59.93 GiB10.63 GiB71.23 GiB0.77 GiB5±8.3%
magnum-v4-72bQ6_K72.7B59.93 GiB10.63 GiB71.23 GiB0.77 GiB5±8.3%
KAT-Dev-72B-ExpQ6_K72.7B59.93 GiB10.63 GiB71.23 GiB0.77 GiB5±8.3%
Chuluun-Qwen2.5-72B-v0.01Q6_K72.7B59.93 GiB10.63 GiB71.23 GiB0.77 GiB5±8.3%
Homer-v1.0-Qwen2.5-72BQ6_K72.7B59.93 GiB10.63 GiB71.23 GiB0.77 GiB5±8.3%
Qwen2.5-VL-72B-InstructQ6_K73.4B59.93 GiB10.63 GiB71.23 GiB0.77 GiB5±8.3%
Tower-Plus-72B-ultra-uncensored-hereticI1-Q6_K72.7B59.93 GiB10.63 GiB71.23 GiB0.77 GiB5±8.3%
Chronos-Platinum-72BQ6_K72.7B59.93 GiB10.63 GiB71.23 GiB0.77 GiB5±8.3%
UI-TARS-72B-DPOQ6_K73.4B59.93 GiB10.63 GiB71.23 GiB0.77 GiB5±8.3%
Laguna-S-2.1MoEQ4_1118B68.96 GiB1.67 GiB71.20 GiB0.80 GiB21±37%
Mixtral-8x22B-Instruct-v0.1MoEQ3_K_M141B63.14 GiB7.44 GiB71.19 GiB0.81 GiB7±37%
Mixtral-8x22B-v0.1MoEQ3_K_M141B63.14 GiB7.44 GiB71.19 GiB0.81 GiB7±37%
Mixtral-8x22B-v0.1MoEQ3_K_M141B63.13 GiB7.44 GiB71.18 GiB0.82 GiB7±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEI1-IQ2_XS229B62.35 GiB8.23 GiB71.12 GiB0.88 GiB13±37%
MiniMax-M2.1MoEI1-IQ2_XS229B62.35 GiB8.23 GiB71.12 GiB0.88 GiB13±37%
MiniMax-M2.5MoEI1-IQ2_XS229B62.35 GiB8.23 GiB71.12 GiB0.88 GiB13±37%
Qwen3.5-122B-A10BMoEQ4_K_S125B69.66 GiB0.80 GiB71.04 GiB0.96 GiB24±37%
Qwen3-235B-A22B-abliteratedMoEI1-IQ2_XS235B64.09 GiB6.24 GiB70.92 GiB1.08 GiB13±37%
Hy3MoEIQ1_S299B59.65 GiB10.63 GiB70.86 GiB1.14 GiB11±37%
Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16MoEQ8_035.1B69.57 GiB0.66 GiB70.79 GiB1.21 GiB25±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q3_K_M139B62.01 GiB8.23 GiB70.78 GiB1.22 GiB12±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-Q3_K_M139B62.01 GiB8.23 GiB70.78 GiB1.22 GiB12±37%
WizardLM-Uncensored-SuperCOT-StoryTelling-30bQ4_K_M32.5B18.27 GiB51.80 GiB70.69 GiB1.31 GiB5±8.3%
Wizard-Vicuna-30B-UncensoredI1-Q4_K_M32.5B18.27 GiB51.80 GiB70.69 GiB1.31 GiB5±8.3%
archangel_sft-kto_llama30bI1-Q4_K_M32.5B18.27 GiB51.80 GiB70.69 GiB1.31 GiB5±8.3%
Qwen3.5-122B-A10B-hereticMoEI1-Q4_K_M123B69.11 GiB0.80 GiB70.49 GiB1.51 GiB25±37%
GLM-Z1-Rumination-32B-0414BF1633.1B61.74 GiB8.10 GiB70.48 GiB1.52 GiB5±8.3%
CallerBF1632.8B61.04 GiB8.50 GiB70.18 GiB1.82 GiB5±8.3%
Dumpling-Qwen2.5-32BBF1632.8B61.04 GiB8.50 GiB70.18 GiB1.82 GiB5±8.3%
OpenThinker-32BF1632.8B61.04 GiB8.50 GiB70.18 GiB1.82 GiB5±8.3%
INTELLECT-2BF1632.8B61.04 GiB8.50 GiB70.18 GiB1.82 GiB5±8.3%
openhands-lm-32b-v0.1BF1632.8B61.04 GiB8.50 GiB70.18 GiB1.82 GiB5±8.3%
LongWriter-Zero-32BBF1632.8B61.04 GiB8.50 GiB70.18 GiB1.82 GiB5±8.3%
OpenCodeReasoning-Nemotron-32BBF1632.8B61.04 GiB8.50 GiB70.18 GiB1.82 GiB5±8.3%
OpenCodeReasoning-Nemotron-32B-IOIBF1632.8B61.04 GiB8.50 GiB70.18 GiB1.82 GiB5±8.3%
Qwen2.5-Coder-32B-Instruct-abliteratedF1632.8B61.04 GiB8.50 GiB70.18 GiB1.82 GiB5±8.3%
QwQ-32B-ArliAI-RpR-v4BF1632.8B61.04 GiB8.50 GiB70.18 GiB1.82 GiB5±8.3%
OpenThinker2-32BBF1632.8B61.04 GiB8.50 GiB70.18 GiB1.82 GiB5±8.3%
Qwen2.5-Coder-32BF1632.8B61.04 GiB8.50 GiB70.18 GiB1.82 GiB5±8.3%
Qwen2.5-32B-InstructF1632.8B61.04 GiB8.50 GiB70.18 GiB1.82 GiB5±8.3%
QwQ-32B-PreviewBF1632.8B61.04 GiB8.50 GiB70.18 GiB1.82 GiB5±8.3%
Qwen2.5-32b-RP-InkF1632.8B61.04 GiB8.50 GiB70.18 GiB1.82 GiB5±8.3%
●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?
2064 of 2118 indexed open-weight models fit a Apple M2 Max at 65,536 context with q8_0 KV cache, the largest being MiniMax-M2.7 at IQ2_XS. That covers text, vision-language, image, video and speech models.
How much usable memory does a Apple M2 Max actually have?
Its nameplate is 96 GB, but about 66.96 GiB is available to a model once driver and compositor overhead is accounted for, and only 72 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.