Apple · apple

Apple M3 Max

Apple M3 Max has 96 GB of unified memory at 307 GB/s — about 66.96 GiB usable after driver and compositor overhead. 2013 of 2118 indexed models fit at 128K context with f16 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
307 GB/s
384-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
vision language 185text 1724image 2audio asr 39audio tts 21embedding 26video 16

What fits at 128K context

largest quantization that fits, per model · 2013 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Qwen3.5-122B-A10BMoEUD-Q4_K_S125B68.39 GiB3.00 GiB71.96 GiB0.04 GiB15±37%
OLMo-2-1124-7B-InstructQ8_07.3B7.23 GiB64.00 GiB71.80 GiB0.20 GiB4±8.3%
GLM-4.6VMoEQ3_K_S108B48.19 GiB23.00 GiB71.77 GiB0.23 GiB5±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ3_K_M124B60.21 GiB11.00 GiB71.76 GiB0.24 GiB8±37%
CodeLlama-70b-Instruct-hfI1-Q3_K_M69.0B30.99 GiB40.00 GiB71.66 GiB0.34 GiB4±8.3%
CodeLlama-70b-Python-hfI1-Q3_K_M69.0B30.99 GiB40.00 GiB71.66 GiB0.34 GiB4±8.3%
Nous-Hermes-Llama2-70bI1-Q3_K_M69.0B30.99 GiB40.00 GiB71.66 GiB0.34 GiB4±8.3%
Midnight-Miqu-70B-v1.5I1-Q3_K_M69.0B30.99 GiB40.00 GiB71.66 GiB0.34 GiB4±8.3%
KafkaLM-70B-German-V0.1Q3_K_M69.0B30.99 GiB40.00 GiB71.66 GiB0.34 GiB4±8.3%
llama2_70b_chat_uncensoredQ3_K_M69.0B30.91 GiB40.00 GiB71.58 GiB0.42 GiB4±8.3%
Xwin-LM-70b-V0.1Q3_K_M69.0B30.91 GiB40.00 GiB71.58 GiB0.42 GiB4±8.3%
Llama-2-70b-chat-hfQ3_K_M69.0B30.91 GiB40.00 GiB71.58 GiB0.42 GiB4±8.3%
CalmeRys-78B-Orpo-v0.1I1-IQ2_S78.0B27.87 GiB43.00 GiB71.55 GiB0.45 GiB4±8.3%
deepseek-llm-67b-chatI1-Q2_K67.4B23.40 GiB47.50 GiB71.55 GiB0.45 GiB4±8.3%
deepseek-llm-67b-baseI1-Q2_K67.4B23.40 GiB47.50 GiB71.55 GiB0.45 GiB4±8.3%
openbuddy-deepseek-67b-v15.3-4kI1-Q2_K67.4B23.40 GiB47.50 GiB71.55 GiB0.45 GiB4±8.3%
GLM-4.5-AirMoEUD-IQ3_XXS110B47.91 GiB23.00 GiB71.49 GiB0.51 GiB5±37%
GLM-4.5VMoEI1-IQ3_M108B47.88 GiB23.00 GiB71.46 GiB0.54 GiB5±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedIQ3_M109B46.87 GiB24.00 GiB71.44 GiB0.56 GiB5±37%
Mistral-Small-4-119B-2603MoEQ4_K_L119B68.02 GiB2.81 GiB71.41 GiB0.59 GiB15±37%
Assistant_Pepe_70BIQ3_M70.6B30.72 GiB40.00 GiB71.40 GiB0.60 GiB4±8.3%
gpt-oss-120b-Uncensored-xCloudMoEI1-Q3_K_M117B66.24 GiB4.53 GiB71.30 GiB0.70 GiB13±37%
gpt-oss-120b-abliteratedMoEI1-Q3_K_M117B66.24 GiB4.53 GiB71.30 GiB0.70 GiB13±37%
Rombo-LLM-V3.0-Qwen-72bI1-IQ3_XS72.7B30.59 GiB40.00 GiB71.27 GiB0.73 GiB4±8.3%
Qwen2.5-72B-Instruct-abliteratedI1-IQ3_XS72.7B30.59 GiB40.00 GiB71.27 GiB0.73 GiB4±8.3%
Qwen2.5-72B-Instruct-abliterated-v2I1-IQ3_XS72.7B30.59 GiB40.00 GiB71.27 GiB0.73 GiB4±8.3%
MiroThinker-v1.0-72BI1-IQ3_XS72.7B30.59 GiB40.00 GiB71.27 GiB0.73 GiB4±8.3%
Malaysian-Qwen2.5-72B-InstructI1-IQ3_XS72.7B30.59 GiB40.00 GiB71.27 GiB0.73 GiB4±8.3%
Qwen2.5-72BI1-IQ3_XS72.7B30.59 GiB40.00 GiB71.27 GiB0.73 GiB4±8.3%
magnum-v4-72bI1-IQ3_XS72.7B30.59 GiB40.00 GiB71.27 GiB0.73 GiB4±8.3%
KAT-Dev-72B-ExpIQ3_XS72.7B30.59 GiB40.00 GiB71.27 GiB0.73 GiB4±8.3%
Tower-Plus-72B-ultra-uncensored-hereticI1-IQ3_XS72.7B30.59 GiB40.00 GiB71.27 GiB0.73 GiB4±8.3%
Qwen2.5-VL-72B-InstructIQ3_XS73.4B30.59 GiB40.00 GiB71.27 GiB0.73 GiB4±8.3%
deepseek-coder-6.7b-instructQ8_06.7B6.67 GiB64.00 GiB71.25 GiB0.75 GiB4±8.3%
deepseek-coder-6.7b-baseQ8_06.7B6.67 GiB64.00 GiB71.25 GiB0.75 GiB4±8.3%
deepseek-coder-6.7B-kexerQ8_06.7B6.67 GiB64.00 GiB71.25 GiB0.75 GiB4±8.3%
Magicoder-S-DS-6.7BQ8_06.7B6.67 GiB64.00 GiB71.25 GiB0.75 GiB4±8.3%
MathCoder2-CodeLlama-7BQ8_06.7B6.67 GiB64.00 GiB71.24 GiB0.76 GiB4±8.3%
CodeLlama-7b-instruct-hfQ8_06.7B6.67 GiB64.00 GiB71.24 GiB0.76 GiB4±8.3%
CodeLlama-7b-hfQ8_06.7B6.67 GiB64.00 GiB71.24 GiB0.76 GiB4±8.3%
WizardLM-7B-UncensoredQ8_06.7B6.67 GiB64.00 GiB71.24 GiB0.76 GiB4±8.3%
Llama-2-7b-chat-hfQ8_06.7B6.67 GiB64.00 GiB71.24 GiB0.76 GiB4±8.3%
Swallow-7b-NVE-instruct-hfQ8_06.7B6.67 GiB64.00 GiB71.24 GiB0.76 GiB4±8.3%
llava-v1.5-7bQ8_06.7B6.67 GiB64.00 GiB71.24 GiB0.76 GiB4±8.3%
Llama-2-7B-32K-InstructQ8_06.7B6.67 GiB64.00 GiB71.24 GiB0.76 GiB4±8.3%
CodeLlama-7b-python-hfQ8_06.7B6.67 GiB64.00 GiB71.24 GiB0.76 GiB4±8.3%
Luna-AI-Llama2-UncensoredQ8_06.7B6.67 GiB64.00 GiB71.24 GiB0.76 GiB4±8.3%
Wizard-Vicuna-7B-UncensoredQ8_06.7B6.67 GiB64.00 GiB71.24 GiB0.76 GiB4±8.3%
llama2_7b_chat_uncensoredQ8_06.7B6.67 GiB64.00 GiB71.24 GiB0.76 GiB4±8.3%
WizardLM-7B-V1.0-UncensoredQ8_06.7B6.67 GiB64.00 GiB71.24 GiB0.76 GiB4±8.3%
Llama-2-7b-hfQ8_06.7B6.67 GiB64.00 GiB71.24 GiB0.76 GiB4±8.3%
pygmalion-2-7bQ8_06.7B6.67 GiB64.00 GiB71.24 GiB0.76 GiB4±8.3%
Devstral-2-123B-Instruct-2512UD-IQ1_S125B26.49 GiB44.00 GiB71.20 GiB0.80 GiB4±8.3%
Behemoth-X-123B-v2IQ1_M123B26.44 GiB44.00 GiB71.14 GiB0.86 GiB4±8.3%
Mistral-Large-Instruct-2411IQ1_M123B26.44 GiB44.00 GiB71.14 GiB0.86 GiB4±8.3%
gpt-oss-20b-hereticMoEIQ4_NL20.9B67.58 GiB3.02 GiB71.13 GiB0.87 GiB9±37%
EXAONE-4.5-33BBF1634.4B61.59 GiB8.84 GiB71.09 GiB0.91 GiB4±8.3%
Laguna-S-2.1MoEQ4_K_S118B64.36 GiB6.14 GiB71.08 GiB0.92 GiB11±37%
IQuest-Coder-V1-40B-InstructI1-Q6_K39.8B30.41 GiB40.00 GiB71.05 GiB0.95 GiB4±8.3%
GLM-4.5-Air-DerestrictedMoEIQ3_XS110B47.35 GiB23.00 GiB70.93 GiB1.07 GiB5±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 M3 Max run?
2013 of 2118 indexed open-weight models fit a Apple M3 Max at 131,072 context with f16 KV cache, the largest being Qwen3.5-122B-A10B at UD-Q4_K_S. That covers text, vision-language, image, video and speech models.
How much usable memory does a Apple M3 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 M3 Max fast for local AI?
Its memory bandwidth is 307 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.