Apple · apple

Apple M3 Max

Apple M3 Max has 48 GB of unified memory at 410 GB/s — about 33.48 GiB usable after driver and compositor overhead. 2001 of 2118 indexed models fit at 64K context with q8_0 KV. Note only 36 GB of its 48 GB is allocatable to the GPU.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
48 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
vision language 180text 1717video 16audio tts 21audio asr 39image 2embedding 26

What fits at 64K context

largest quantization that fits, per model · 2001 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
EXAONE-4.5-33BQ8_034.4B32.73 GiB2.57 GiB35.95 GiB0.05 GiB9±8.3%
MythoMax-L2-Kimiko-v2-13bQ5_K_M13.0B8.78 GiB26.56 GiB35.93 GiB0.07 GiB9±8.3%
MythoMax-L2-13bI1-Q5_K_M13.0B8.78 GiB26.56 GiB35.93 GiB0.07 GiB9±8.3%
Qwen3-VL-8B-Instruct-HereticF168.8B30.53 GiB4.78 GiB35.89 GiB0.11 GiB9±8.3%
MiniCPM-V-4_5F168.7B30.52 GiB4.78 GiB35.88 GiB0.12 GiB9±8.3%
c4ai-command-r-plus-08-2024IQ2_XXS104B26.65 GiB8.50 GiB35.88 GiB0.12 GiB9±8.3%
Maenad-70BI1-Q2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-ReasonerI1-Q2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
calme-2.4-llama3-70bQ2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
calme-2.2-llama3-70bQ2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
Rombos-LLM-70b-Llama-3.3I1-Q2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
L3.3-Electra-R1-70bI1-Q2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
L3.3-70B-Magnum-v4-SEQ2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
Latxa-Llama-3.1-70B-Instruct-v2I1-Q2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
Llama-3.3_70_b_uncensored_continuedI1-Q2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
Llama-3.3-70B-Instruct-abliteratedI1-Q2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
Strawberrylemonade-L3-70B-v1.2Q2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
grok-oss-Revenant-70BI1-Q2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
Llama-3.1-Nemotron-70B-Instruct-HFI1-Q2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
L3.3-70B-Euryale-v2.3I1-Q2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
Hermes-4-70B-hereticI1-Q2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
Hermes-4-70BQ2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
Llama-3.3-70B-InstructQ2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
Llama-3.1-70BQ2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
Anubis-70B-v1.2Q2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
Golem-70B-v1bI1-Q2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
DeepSeek-R1-Distill-Llama-70B-abliteratedI1-Q2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
DeepSeek-R1-Distill-Llama-70B-hereticI1-Q2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
DeepSeek-R1-Distill-Llama-70BQ2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
llama-3-firefunction-v2Q2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
Legion-V2.1-LLaMa-70BI1-Q2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
Assistant_Pepe_70BI1-Q2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
Tess-R1-Limerick-Llama-3.1-70BQ2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
SEMIKONG-70BQ2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
functionary-medium-v3.2KV unresolvedQ2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
Llama-3.1-WhiteRabbitNeo-2-70BQ2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
Infinity-Instruct-7M-Gen-Llama3_1-70BI1-Q2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
New-Dawn-Llama-3-70B-32K-v1.0I1-Q2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
Meta-Llama-3-70B-Instruct-abliterated-v3.5I1-Q2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
Athene-70BQ2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
Hermes-3-Llama-3.1-70BQ2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
L3.3-70B-Magnum-DiamondQ2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
Meta-Llama-3-70B-InstructQ2_K70.6B24.56 GiB10.63 GiB35.86 GiB0.14 GiB9±8.3%
xLAM-8x7b-rMoEQ5_K_L46.7B31.02 GiB4.25 GiB35.86 GiB0.14 GiB14±37%
NSFW_13B_sftQ5_K_S13.3B8.70 GiB26.56 GiB35.86 GiB0.14 GiB9±8.3%
Mistral-Small-4-119B-2603MoEIQ2_XS119B34.48 GiB0.75 GiB35.81 GiB0.19 GiB40±37%
Qwen3.6-35B-A3B-REAM-192-hereticMoEQ5_K_M27.0B34.58 GiB0.66 GiB35.80 GiB0.20 GiB38±37%
dolphin-2.6-mixtral-8x7bMoEI1-Q5_K_M46.7B30.95 GiB4.25 GiB35.78 GiB0.22 GiB14±37%
Nous-Hermes-2-Mixtral-8x7B-DPOMoEQ5_K_M46.7B30.95 GiB4.25 GiB35.78 GiB0.22 GiB14±37%
Mixtral-8x7B-Instruct-v0.1MoEQ5_K_M46.7B30.95 GiB4.25 GiB35.78 GiB0.22 GiB14±37%
dolphin-2.5-mixtral-8x7bMoEQ5_K_M46.7B30.95 GiB4.25 GiB35.78 GiB0.22 GiB14±37%
Mixtral-8x7B-v0.1MoEQ5_K_M46.7B30.95 GiB4.25 GiB35.78 GiB0.22 GiB14±37%
Open_Gpt4_8x7B_v0.2MoEQ5_K_M46.7B30.95 GiB4.25 GiB35.78 GiB0.22 GiB14±37%
CalmeRys-78B-Orpo-v0.1I1-IQ1_M78.0B23.68 GiB11.42 GiB35.78 GiB0.22 GiB9±8.3%
calme-2.3-rys-78bIQ1_M78.0B23.68 GiB11.42 GiB35.78 GiB0.22 GiB9±8.3%
codellama-13b-oasst-sft-v10Q5_K_M13.0B8.60 GiB26.56 GiB35.75 GiB0.25 GiB9±8.3%
chronos-hermes-13b-v2Q5_K_M13.0B8.60 GiB26.56 GiB35.75 GiB0.25 GiB9±8.3%
WhiteRabbitNeo-13B-v1Q5_K_M13.0B8.60 GiB26.56 GiB35.75 GiB0.25 GiB9±8.3%
CodeLlama-13b-Instruct-hfQ5_K_M13.0B8.60 GiB26.56 GiB35.75 GiB0.25 GiB9±8.3%
Orca-2-13b-Alpaca-UncensoredI1-Q5_K_M13.0B8.60 GiB26.56 GiB35.75 GiB0.25 GiB9±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 M3 Max run?
2001 of 2118 indexed open-weight models fit a Apple M3 Max at 65,536 context with q8_0 KV cache, the largest being EXAONE-4.5-33B at Q8_0. That covers text, vision-language, image, video and speech models.
How much usable memory does a Apple M3 Max actually have?
Its nameplate is 48 GB, but about 33.48 GiB is available to a model once driver and compositor overhead is accounted for, and only 36 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 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.