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. 2030 of 2118 indexed models fit at 4K context with q4_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
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 1743vision language 183video 16audio tts 21image 2embedding 26audio asr 39

What fits at 4K context

largest quantization that fits, per model · 2030 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Huihui-Qwen3-Coder-Next-abliteratedMoEQ3_K_L79.7B35.42 GiB0.03 GiB35.98 GiB0.02 GiB60±37%
Assistant_Pepe_70BQ3_K_L70.6B34.94 GiB0.35 GiB35.97 GiB0.03 GiB12±8.3%
gemma-4-31B-it-abliteratedQ4_K_M31.3B34.81 GiB0.51 GiB35.94 GiB0.06 GiB12±8.3%
GLM-4.6VMoEIQ2_XXS108B35.11 GiB0.20 GiB35.89 GiB0.11 GiB45±37%
Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16MoEIQ4_XS35.1B35.30 GiB0.02 GiB35.88 GiB0.12 GiB56±37%
Qwen3-Coder-Next-REAMMoEI1-Q4_160.3B35.30 GiB0.03 GiB35.87 GiB0.13 GiB58±37%
HarmonicHarlequin_v5-20BQ8_033.3B32.97 GiB2.29 GiB35.85 GiB0.15 GiB12±8.3%
Devstral-2-123B-Instruct-2512IQ2_XS125B34.75 GiB0.39 GiB35.85 GiB0.15 GiB12±8.3%
Mistral-Medium-3.5-128BI1-IQ2_XS128B34.75 GiB0.39 GiB35.85 GiB0.15 GiB12±8.3%
XORTRON-NXTXPRTXXLI1-IQ2_XS128B34.75 GiB0.39 GiB35.85 GiB0.15 GiB12±8.3%
Qwen3-53B-A3B-2507-THINKING-TOTAL-RECALL-v2-MASTER-CODERMoEI1-Q5_K_M53.0B35.08 GiB0.18 GiB35.80 GiB0.20 GiB41±37%
Salience-1.5-ProMoEQ8_036.0B35.22 GiB0.02 GiB35.79 GiB0.21 GiB56±37%
Qwable-v1MoEQ8_036.0B35.22 GiB0.02 GiB35.79 GiB0.21 GiB56±37%
T-SearchMoEQ8_036.0B35.22 GiB0.02 GiB35.79 GiB0.21 GiB56±37%
Qwen3.5-35B-A3BMoEQ8_036.0B35.22 GiB0.02 GiB35.79 GiB0.21 GiB56±37%
Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-PreservedMoEQ8_035.1B35.21 GiB0.02 GiB35.78 GiB0.22 GiB56±37%
Qwen3.6-35B-A3B-Fable-5-DistillMoEQ8_036.0B35.21 GiB0.02 GiB35.78 GiB0.22 GiB56±37%
Qwable-v2MoEQ8_036.0B35.21 GiB0.02 GiB35.78 GiB0.22 GiB56±37%
Qwen3.6-35B-A3B-YOYO-V2MoEQ8_036.0B35.21 GiB0.02 GiB35.78 GiB0.22 GiB56±37%
Ornith-1.0-35B-FP8-BLOCK-MTPMoEQ8_035.5B35.21 GiB0.02 GiB35.78 GiB0.22 GiB56±37%
fable-coder-35B-A3BMoEQ8_036.0B35.21 GiB0.02 GiB35.78 GiB0.22 GiB56±37%
PINQWEN-3.6-35B-CLEAN-BF16MoEQ8_036.0B35.21 GiB0.02 GiB35.78 GiB0.22 GiB56±37%
UniMath-35B-A3BMoEQ8_036.0B35.21 GiB0.02 GiB35.78 GiB0.22 GiB56±37%
Ornith-1.0-35B-Heretic-MTPMoEQ8_0—35.21 GiB0.02 GiB35.78 GiB0.22 GiB56±37%
Fawen-1.0-35BMoEQ8_036.0B35.21 GiB0.02 GiB35.78 GiB0.22 GiB56±37%
Qwopus3.6-35B-A3B-v1MoEQ8_036.0B35.21 GiB0.02 GiB35.78 GiB0.22 GiB56±37%
CyberStrike-OffSec-35BMoEQ8_035.1B35.21 GiB0.02 GiB35.78 GiB0.22 GiB56±37%
Qwen3.6-35B-A3BMoEQ8_036.0B35.21 GiB0.02 GiB35.78 GiB0.22 GiB56±37%
Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-DistilledMoEQ8_036.0B35.21 GiB0.02 GiB35.78 GiB0.22 GiB56±37%
Qwen3.5-35B-A3B-uncensored-heretic-v2-Native-MTP-PreservedMoEQ8_035.1B35.21 GiB0.02 GiB35.78 GiB0.22 GiB56±37%
CodeLlama-70b-Python-hfIQ4_XS69.0B34.64 GiB0.35 GiB35.67 GiB0.33 GiB12±8.3%
Midnight-Miqu-70B-v1.5IQ4_XS69.0B34.64 GiB0.35 GiB35.67 GiB0.33 GiB12±8.3%
Qwen2.5-Coder-32B-InstructQ4_032.8B34.72 GiB0.28 GiB35.65 GiB0.35 GiB12±8.3%
Meta-Llama-3-70B-InstructQ3_K_L70.6B34.60 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedUD-IQ2_XXS109B34.83 GiB0.21 GiB35.62 GiB0.38 GiB45±37%
Maenad-70BI1-Q3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-ReasonerI1-Q3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
calme-2.4-llama3-70bQ3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
calme-2.2-llama3-70bQ3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
Rombos-LLM-70b-Llama-3.3I1-Q3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
L3.3-Electra-R1-70bI1-Q3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
L3.3-70B-Magnum-v4-SEQ3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
Latxa-Llama-3.1-70B-Instruct-v2I1-Q3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
Llama-3.3_70_b_uncensored_continuedI1-Q3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
Llama-3.3-70B-Instruct-abliteratedI1-Q3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
Strawberrylemonade-L3-70B-v1.2Q3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
grok-oss-Revenant-70BI1-Q3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
Llama-3.1-Nemotron-70B-Instruct-HFI1-Q3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
L3.3-70B-Euryale-v2.3I1-Q3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
Hermes-4-70B-hereticI1-Q3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
Llama-3.3-70B-InstructQ3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
Llama-3.1-70BQ3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
Hermes-3-Llama-3.1-70BQ3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
Anubis-70B-v1.2Q3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
Hermes-4-70BQ3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
Golem-70B-v1bI1-Q3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
DeepSeek-R1-Distill-Llama-70B-abliteratedI1-Q3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
DeepSeek-R1-Distill-Llama-70B-hereticI1-Q3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
llama-3-firefunction-v2Q3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±8.3%
DeepSeek-R1-Distill-Llama-70BQ3_K_L70.6B34.59 GiB0.35 GiB35.62 GiB0.38 GiB12±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?
2030 of 2118 indexed open-weight models fit a Apple M4 Max at 4,096 context with q4_0 KV cache, the largest being Huihui-Qwen3-Coder-Next-abliterated at Q3_K_L. 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.
Apple M4 Max — what AI models can it run locally? — ossmodeldb