Apple · apple

Apple M4 Max

Apple M4 Max has 36 GB of unified memory at 410 GB/s — about 25.11 GiB usable after driver and compositor overhead. 2011 of 2118 indexed models fit at 4K context with q8_0 KV. Note only 27 GB of its 36 GB is allocatable to the GPU.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
36 GB
LPDDR5X-8533
Bandwidth
410 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
text 1727vision language 180audio tts 21audio asr 39video 16image 2embedding 26

What fits at 4K context

largest quantization that fits, per model · 2011 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
gemma-4-A4B-98e-v6-coder-itMoEQ5_K_M20.5B26.21 GiB0.24 GiB26.98 GiB0.02 GiB12±8.3%
NSFW_13B_sftF1613.3B24.71 GiB1.66 GiB26.96 GiB0.04 GiB12±8.3%
Mistral-Small-Instruct-2409IQ2_XS22.2B25.88 GiB0.46 GiB26.96 GiB0.04 GiB12±8.3%
Qwen3.5-14B-A3B-Claude-4.6-Opus-Reasoning-Distilled-reapMoEBF1614.1B26.36 GiB0.04 GiB26.95 GiB0.05 GiB40±37%
Apertus-70B-Instruct-2509IQ3_XXS70.6B25.53 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
Maenad-70BI1-IQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-ReasonerI1-IQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
Rombos-LLM-70b-Llama-3.3I1-IQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
L3.3-Electra-R1-70bI1-IQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
L3.3-70B-Magnum-v4-SEIQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
Latxa-Llama-3.1-70B-Instruct-v2I1-IQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
Llama-3.3_70_b_uncensored_continuedI1-IQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
Llama-3.3-70B-Instruct-abliteratedI1-IQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
grok-oss-Revenant-70BI1-IQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
Llama-3.1-Nemotron-70B-Instruct-HFI1-IQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
L3.3-70B-Euryale-v2.3I1-IQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
Hermes-3-Llama-3.1-70BIQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
Hermes-4-70B-hereticI1-IQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
Llama-3.3-70B-InstructIQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
Llama-3.1-70BIQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
Anubis-70B-v1.2IQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
Hermes-4-70BIQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
Golem-70B-v1bI1-IQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
DeepSeek-R1-Distill-Llama-70B-abliteratedI1-IQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
DeepSeek-R1-Distill-Llama-70B-hereticI1-IQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
DeepSeek-R1-Distill-Llama-70BIQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
Legion-V2.1-LLaMa-70BI1-IQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
Assistant_Pepe_70BI1-IQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
Tess-R1-Limerick-Llama-3.1-70BIQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
SEMIKONG-70BIQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
functionary-medium-v3.2KV unresolvedIQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
Llama-3.1-WhiteRabbitNeo-2-70BIQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
Infinity-Instruct-7M-Gen-Llama3_1-70BI1-IQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
New-Dawn-Llama-3-70B-32K-v1.0I1-IQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
Meta-Llama-3-70B-Instruct-abliterated-v3.5I1-IQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
Athene-70BIQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
L3.3-70B-Magnum-DiamondIQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
Meta-Llama-3-70B-InstructIQ3_XXS70.6B25.58 GiB0.66 GiB26.92 GiB0.08 GiB12±8.3%
Hypernova-60B-2605MoEI1-IQ2_XS58.7B26.29 GiB0.08 GiB26.91 GiB0.09 GiB49±37%
GLM-Z1-Rumination-32B-0414Q6_K_L33.1B25.75 GiB0.51 GiB26.90 GiB0.10 GiB12±8.3%
HarmonicHarlequin_v5-20BI1-Q5_K_M33.3B21.98 GiB4.32 GiB26.89 GiB0.11 GiB12±8.3%
Huihui-Qwen3-Coder-Next-abliteratedMoEQ2_K_L79.7B26.29 GiB0.05 GiB26.88 GiB0.12 GiB60±37%
IQuest-Coder-V1-40B-InstructI1-Q5_K_S39.8B25.53 GiB0.66 GiB26.84 GiB0.16 GiB12±8.3%
WizardCoder-Python-34B-V1.0I1-Q6_K33.7B25.78 GiB0.40 GiB26.83 GiB0.17 GiB12±8.3%
Phind-CodeLlama-34B-Python-v1I1-Q6_K33.7B25.78 GiB0.40 GiB26.83 GiB0.17 GiB12±8.3%
Phind-CodeLlama-34B-v2I1-Q6_K33.7B25.78 GiB0.40 GiB26.83 GiB0.17 GiB12±8.3%
CodeLlama-34b-instruct-hfQ6_K33.7B25.78 GiB0.40 GiB26.83 GiB0.17 GiB12±8.3%
WizardLM-1.0-Uncensored-CodeLlama-34bQ6_K33.7B25.78 GiB0.40 GiB26.83 GiB0.17 GiB12±8.3%
Magistral-Small-2509-VisionQ6_K_L24.0B25.83 GiB0.33 GiB26.83 GiB0.17 GiB12±8.3%
CalmeRys-78B-Orpo-v0.1I1-IQ2_XXS78.0B25.43 GiB0.71 GiB26.82 GiB0.18 GiB12±8.3%
calme-2.3-rys-78bIQ2_XXS78.0B25.43 GiB0.71 GiB26.82 GiB0.18 GiB12±8.3%
Mistral-MOE-4X7B-Dark-MultiVerse-Uncensored-Enhanced32-24BMoEQ2_K24.2B25.95 GiB0.27 GiB26.80 GiB0.20 GiB7±37%
Gemma4-Gutenberg-31BQ6_K_L31.3B25.21 GiB0.95 GiB26.79 GiB0.21 GiB12±8.3%
gemma-4-31B-itQ6_K_L31.3B25.21 GiB0.95 GiB26.79 GiB0.21 GiB12±8.3%
Gemma4-Gutenberg-31B-HereticQ6_K_L31.3B25.21 GiB0.95 GiB26.79 GiB0.21 GiB12±8.3%
Equinox-31BQ6_K_L31.3B25.21 GiB0.95 GiB26.79 GiB0.21 GiB12±8.3%
gemma-4-31B-it-SDFT-Heretic-RPQ6_K_L30.7B25.21 GiB0.95 GiB26.79 GiB0.21 GiB12±8.3%
Qwen3.5-40B-RoughHouse-Claude-4.6-Opus-Polar-Deckard-Uncensored-Heretic-ThinkingI1-Q5_K_M39.5B25.97 GiB0.20 GiB26.78 GiB0.22 GiB12±8.3%
Qwen3.5-40B-Claude-4.5-Opus-High-Reasoning-Thinking-uncensored-hereticQ5_K_M39.5B25.97 GiB0.20 GiB26.78 GiB0.22 GiB12±8.3%
Devstral-2-123B-Instruct-2512IQ1_S125B25.33 GiB0.73 GiB26.76 GiB0.24 GiB13±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 M4 Max run?
2011 of 2118 indexed open-weight models fit a Apple M4 Max at 4,096 context with q8_0 KV cache, the largest being gemma-4-A4B-98e-v6-coder-it at Q5_K_M. That covers text, vision-language, image, video and speech models.
How much usable memory does a Apple M4 Max actually have?
Its nameplate is 36 GB, but about 25.11 GiB is available to a model once driver and compositor overhead is accounted for, and only 27 GB of the pool can be allocated to the GPU at all.
Is a Apple M4 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.