Apple · apple

Apple M5 Max

Apple M5 Max has 64 GB of unified memory at 614 GB/s — about 44.64 GiB usable after driver and compositor overhead. 1994 of 2118 indexed models fit at 64K 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
LPDDR5X-9600
Bandwidth
614 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 1707image 2vision language 183audio tts 21audio asr 39video 16embedding 26

What fits at 64K context

largest quantization that fits, per model · 1994 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Maenad-70BI1-IQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-ReasonerI1-IQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
Rombos-LLM-70b-Llama-3.3I1-IQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
L3.3-Electra-R1-70bI1-IQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
Latxa-Llama-3.1-70B-Instruct-v2I1-IQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
Llama-3.3_70_b_uncensored_continuedI1-IQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
Llama-3.3-70B-Instruct-abliteratedI1-IQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
grok-oss-Revenant-70BI1-IQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
Llama-3.1-Nemotron-70B-Instruct-HFI1-IQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
L3.3-70B-Euryale-v2.3I1-IQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
Hermes-3-Llama-3.1-70BIQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
Hermes-4-70B-hereticI1-IQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
Llama-3.3-70B-InstructIQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
Llama-3.1-70BIQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
Anubis-70B-v1.2IQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
Hermes-4-70BIQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
Golem-70B-v1bI1-IQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
DeepSeek-R1-Distill-Llama-70B-abliteratedI1-IQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
DeepSeek-R1-Distill-Llama-70B-hereticI1-IQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
llama-3-firefunction-v2IQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
Legion-V2.1-LLaMa-70BI1-IQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
Assistant_Pepe_70BI1-IQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
Tess-R1-Limerick-Llama-3.1-70BIQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
Infinity-Instruct-7M-Gen-Llama3_1-70BI1-IQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
New-Dawn-Llama-3-70B-32K-v1.0I1-IQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
Meta-Llama-3-70B-Instruct-abliterated-v3.5I1-IQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
Athene-70BIQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
L3.3-70B-Magnum-DiamondIQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
Meta-Llama-3-70B-InstructIQ3_XS70.6B27.29 GiB20.00 GiB47.97 GiB0.03 GiB11±8.3%
Nous-Hermes-Llama2-70bQ2_K69.0B27.27 GiB20.00 GiB47.94 GiB0.06 GiB11±8.3%
llama2_70b_chat_uncensoredQ2_K69.0B27.27 GiB20.00 GiB47.94 GiB0.06 GiB11±8.3%
Xwin-LM-70b-V0.1Q2_K69.0B27.27 GiB20.00 GiB47.94 GiB0.06 GiB11±8.3%
Llama-2-70b-chat-hfQ2_K69.0B27.27 GiB20.00 GiB47.94 GiB0.06 GiB11±8.3%
GLM-4.5-AirMoEUD-IQ1_S110B35.86 GiB11.50 GiB47.94 GiB0.06 GiB17±37%
reka-flash-3BF1620.9B38.94 GiB8.25 GiB47.82 GiB0.18 GiB11±8.3%
reka-flash-3.1BF1620.9B38.94 GiB8.25 GiB47.82 GiB0.18 GiB11±8.3%
deepseek-llm-67b-chatI1-Q2_K67.4B23.40 GiB23.75 GiB47.80 GiB0.20 GiB11±8.3%
deepseek-llm-67b-baseI1-Q2_K67.4B23.40 GiB23.75 GiB47.80 GiB0.20 GiB11±8.3%
openbuddy-deepseek-67b-v15.3-4kI1-Q2_K67.4B23.40 GiB23.75 GiB47.80 GiB0.20 GiB11±8.3%
c4ai-command-r-plus-08-2024IQ2_S104B31.04 GiB16.00 GiB47.77 GiB0.23 GiB11±8.3%
GLM-4.5VMoEI1-IQ2_XXS108B35.61 GiB11.50 GiB47.68 GiB0.32 GiB17±37%
glm-4-9b-chat-1mQ5_K_L9.5B7.07 GiB40.00 GiB47.67 GiB0.33 GiB11±8.3%
Mixtral_34Bx2_MoE_60BMoEQ4_K_M60.8B32.03 GiB15.00 GiB47.66 GiB0.34 GiB6±37%
HunyuanImage-2.1Q5_017.5B47.04 GiB0.00 GiB47.64 GiB0.36 GiB11±8.3%
CalmeRys-78B-Orpo-v0.1I1-IQ2_XXS78.0B25.43 GiB21.50 GiB47.61 GiB0.39 GiB11±8.3%
calme-2.3-rys-78bIQ2_XXS78.0B25.43 GiB21.50 GiB47.61 GiB0.39 GiB11±8.3%
glm-4-9b-chat-abliteratedQ5_K_L9.4B7.01 GiB40.00 GiB47.60 GiB0.40 GiB11±8.3%
glm-4-9b-chatQ5_K_L9.4B7.01 GiB40.00 GiB47.60 GiB0.40 GiB11±8.3%
Huihui-Qwen3-Coder-Next-abliteratedMoEQ4_K_L79.7B45.44 GiB1.50 GiB47.48 GiB0.52 GiB44±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedUD-IQ2_XXS109B34.83 GiB12.00 GiB47.41 GiB0.59 GiB17±37%
Hunyuan-A13B-InstructMoEQ3_K_L80.4B38.84 GiB8.00 GiB47.39 GiB0.61 GiB11±8.3%
Ace-Step1.5Q8_0160M43.77 GiB3.04 GiB47.35 GiB0.65 GiB11±8.3%
GLM-4.5-Air-DerestrictedMoEIQ1_M110B35.26 GiB11.50 GiB47.34 GiB0.66 GiB17±37%
codegeex4-all-9bQ5_K_M9.4B6.65 GiB40.00 GiB47.25 GiB0.75 GiB11±8.3%
Qwen3-Coder-Next-Opus-4.6-Reasoning-DistilledMoEQ4_K_M—45.16 GiB1.50 GiB47.20 GiB0.80 GiB44±37%
GLM-4.6VMoEIQ2_XXS108B35.11 GiB11.50 GiB47.18 GiB0.82 GiB17±37%
Apertus-70B-Instruct-2509Q2_K_L70.6B26.38 GiB20.00 GiB47.11 GiB0.89 GiB11±8.3%
CodeLlama-70b-Instruct-hfI1-IQ3_XS69.0B26.37 GiB20.00 GiB47.04 GiB0.96 GiB11±8.3%
CodeLlama-70b-Python-hfI1-IQ3_XS69.0B26.37 GiB20.00 GiB47.04 GiB0.96 GiB11±8.3%
Midnight-Miqu-70B-v1.5I1-IQ3_XS69.0B26.37 GiB20.00 GiB47.04 GiB0.96 GiB11±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 M5 Max run?
1994 of 2118 indexed open-weight models fit a Apple M5 Max at 65,536 context with f16 KV cache, the largest being Maenad-70B at I1-IQ3_XS. That covers text, vision-language, image, video and speech models.
How much usable memory does a Apple M5 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 M5 Max fast for local AI?
Its memory bandwidth is 614 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.