Apple · apple

Apple M4 Max

Apple M4 Max has 48 GB of unified memory at 546 GB/s — about 33.48 GiB usable after driver and compositor overhead. 2002 of 2118 indexed models fit at 32K context with f16 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
LPDDR5X-8533
Bandwidth
546 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 1718vision language 180video 16audio tts 21audio asr 39embedding 26image 2

What fits at 32K context

largest quantization that fits, per model · 2002 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Laguna-S-2.1MoEIQ2_S118B33.77 GiB1.64 GiB35.98 GiB0.02 GiB41±37%
Hunyuan-A13B-InstructMoEIQ3_XS80.4B31.41 GiB4.00 GiB35.96 GiB0.04 GiB12±8.3%
Noromaid-v0.4-Mixtral-Instruct-8x7b-ZlossMoEQ5_K_M46.7B31.34 GiB4.00 GiB35.92 GiB0.08 GiB18±37%
deepseek-llm-67b-chatI1-Q2_K67.4B23.40 GiB11.88 GiB35.92 GiB0.08 GiB12±8.3%
deepseek-llm-67b-baseI1-Q2_K67.4B23.40 GiB11.88 GiB35.92 GiB0.08 GiB12±8.3%
openbuddy-deepseek-67b-v15.3-4kI1-Q2_K67.4B23.40 GiB11.88 GiB35.92 GiB0.08 GiB12±8.3%
Behemoth-X-123B-v2IQ1_S123B24.18 GiB11.00 GiB35.88 GiB0.12 GiB12±8.3%
Rombo-LLM-V3.0-Qwen-72bI1-IQ2_XS72.7B25.20 GiB10.00 GiB35.88 GiB0.12 GiB12±8.3%
Qwen2.5-72B-Instruct-abliteratedI1-IQ2_XS72.7B25.20 GiB10.00 GiB35.88 GiB0.12 GiB12±8.3%
Qwen2.5-72B-Instruct-abliterated-v2I1-IQ2_XS72.7B25.20 GiB10.00 GiB35.88 GiB0.12 GiB12±8.3%
HuatuoGPT-o1-72BIQ2_XS72.7B25.20 GiB10.00 GiB35.88 GiB0.12 GiB12±8.3%
MiroThinker-v1.0-72BI1-IQ2_XS72.7B25.20 GiB10.00 GiB35.88 GiB0.12 GiB12±8.3%
EVA-Qwen2.5-72B-v0.2IQ2_XS72.7B25.20 GiB10.00 GiB35.88 GiB0.12 GiB12±8.3%
Qwen2.5-Math-72B-InstructIQ2_XS72.7B25.20 GiB10.00 GiB35.88 GiB0.12 GiB12±8.3%
Qwen2.5-72B-InstructIQ2_XS72.7B25.20 GiB10.00 GiB35.88 GiB0.12 GiB12±8.3%
Malaysian-Qwen2.5-72B-InstructI1-IQ2_XS72.7B25.20 GiB10.00 GiB35.88 GiB0.12 GiB12±8.3%
Qwen2.5-72BI1-IQ2_XS72.7B25.20 GiB10.00 GiB35.88 GiB0.12 GiB12±8.3%
magnum-v4-72bI1-IQ2_XS72.7B25.20 GiB10.00 GiB35.88 GiB0.12 GiB12±8.3%
KAT-Dev-72B-ExpIQ2_XS72.7B25.20 GiB10.00 GiB35.88 GiB0.12 GiB12±8.3%
Homer-v1.0-Qwen2.5-72BIQ2_XS72.7B25.20 GiB10.00 GiB35.88 GiB0.12 GiB12±8.3%
Tower-Plus-72B-ultra-uncensored-hereticI1-IQ2_XS72.7B25.20 GiB10.00 GiB35.88 GiB0.12 GiB12±8.3%
Qwen2.5-VL-72B-InstructIQ2_XS73.4B25.20 GiB10.00 GiB35.88 GiB0.12 GiB12±8.3%
Chronos-Platinum-72BIQ2_XS72.7B25.20 GiB10.00 GiB35.88 GiB0.12 GiB12±8.3%
UI-TARS-72B-DPOIQ2_XS73.4B25.20 GiB10.00 GiB35.88 GiB0.12 GiB12±8.3%
Qwen2.5-Coder-14B-InstructQ8_014.8B29.25 GiB6.00 GiB35.84 GiB0.16 GiB12±8.3%
Assistant_Pepe_70BIQ2_M70.6B25.10 GiB10.00 GiB35.78 GiB0.22 GiB12±8.3%
Qwen3.6-35B-A3B-REAM-192-hereticMoEQ5_K_M27.0B34.58 GiB0.63 GiB35.76 GiB0.24 GiB47±37%
Mistral-Small-4-119B-2603MoEIQ2_XS119B34.48 GiB0.70 GiB35.76 GiB0.24 GiB49±37%
Qwen3-53B-A3B-2507-THINKING-TOTAL-RECALL-v2-MASTER-CODERMoEI1-Q4_K_M53.0B29.94 GiB5.25 GiB35.73 GiB0.27 GiB25±37%
MythoMax-L2-Kimiko-v2-13bQ6_K13.0B10.13 GiB25.00 GiB35.72 GiB0.28 GiB12±8.3%
MythoMax-L2-13bI1-Q6_K13.0B10.13 GiB25.00 GiB35.72 GiB0.28 GiB12±8.3%
xLAM-8x7b-rMoEQ5_K_L46.7B31.02 GiB4.00 GiB35.61 GiB0.39 GiB18±37%
Qwen3-VL-8B-Instruct-HereticF168.8B30.53 GiB4.50 GiB35.61 GiB0.39 GiB12±8.3%
MiniCPM-V-4_5F168.7B30.52 GiB4.50 GiB35.60 GiB0.40 GiB12±8.3%
Darwin-35B-A3B-OpusMoEQ8_036.0B34.38 GiB0.63 GiB35.56 GiB0.44 GiB50±37%
Aurora-Code-1MoEQ8_034.7B34.38 GiB0.63 GiB35.56 GiB0.44 GiB50±37%
grug-35b-v2MoEQ8_035.1B34.38 GiB0.63 GiB35.56 GiB0.44 GiB50±37%
grug-35bMoEQ8_035.1B34.38 GiB0.63 GiB35.56 GiB0.44 GiB50±37%
WorldSim-Opus-3.6-35B-A3BMoEQ8_035.1B34.38 GiB0.63 GiB35.56 GiB0.44 GiB50±37%
Qwen3.6-35B-A3B-AnkoMoEQ8_035.1B34.38 GiB0.63 GiB35.56 GiB0.44 GiB50±37%
KAT-Coder-V2.5-DevMoEQ8_034.7B34.38 GiB0.63 GiB35.56 GiB0.44 GiB50±37%
Ornith-1.0-35BMoEQ8_034.7B34.38 GiB0.63 GiB35.56 GiB0.44 GiB50±37%
Nex-N2-miniMoEQ8_035.1B34.38 GiB0.63 GiB35.56 GiB0.44 GiB50±37%
14BQ5_114.2B9.95 GiB25.00 GiB35.55 GiB0.45 GiB12±8.3%
Qwen3.6-35B-A3B-uncensored-hereticMoEQ8_035.1B34.37 GiB0.63 GiB35.55 GiB0.45 GiB50±37%
Qwen35B-Agent-R2-AbliteratedMoEQ8_034.7B34.37 GiB0.63 GiB35.55 GiB0.45 GiB50±37%
spoomplesmaxx-flash-35B-A3MoEQ8_035.1B34.37 GiB0.63 GiB35.55 GiB0.45 GiB50±37%
Holo-3.1-35B-A3BMoEQ8_035.1B34.37 GiB0.63 GiB35.55 GiB0.45 GiB50±37%
Qwen-AgentWorld-35B-A3BMoEQ8_034.7B34.37 GiB0.63 GiB35.55 GiB0.45 GiB50±37%
Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-DistilledMoEQ8_036.0B34.37 GiB0.63 GiB35.55 GiB0.45 GiB50±37%
Qwen3.6-35B-A3B-abliterated-MAXMoEQ8_035.1B34.37 GiB0.63 GiB35.55 GiB0.45 GiB50±37%
Qwen3.6-35B-A3B-StyleTuneMoEQ8_035.1B34.37 GiB0.63 GiB35.55 GiB0.45 GiB50±37%
Huihui-Qwen3.5-35B-A3B-abliteratedMoEQ8_036.0B34.37 GiB0.63 GiB35.55 GiB0.45 GiB50±37%
Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-DistilledMoEQ8_036.0B34.37 GiB0.63 GiB35.55 GiB0.45 GiB50±37%
Qwen3.6-35B-A3B-abliteratedMoEQ8_035.1B34.37 GiB0.63 GiB35.55 GiB0.45 GiB50±37%
Qwen3.6-35B-A3B-hereticMoEQ8_035.1B34.37 GiB0.63 GiB35.55 GiB0.45 GiB50±37%
Infinity-Parser2-ProMoEQ8_035.1B34.37 GiB0.63 GiB35.55 GiB0.45 GiB50±37%
Qwen3.5-35B-A3B-BaseMoEQ8_036.0B34.37 GiB0.63 GiB35.55 GiB0.45 GiB50±37%
Huihui-Qwen3.6-35B-A3B-abliteratedMoEQ8_036.0B34.37 GiB0.63 GiB35.55 GiB0.45 GiB50±37%
Ornith-1.0-35B-uncensored-hereticMoEQ8_035.1B34.37 GiB0.63 GiB35.55 GiB0.45 GiB50±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 M4 Max run?
2002 of 2118 indexed open-weight models fit a Apple M4 Max at 32,768 context with f16 KV cache, the largest being Laguna-S-2.1 at IQ2_S. That covers text, vision-language, image, video and speech models.
How much usable memory does a Apple M4 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 M4 Max fast for local AI?
Its memory bandwidth is 546 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.