Apple · apple

Apple M4 Pro

Apple M4 Pro has 64 GB of unified memory at 273 GB/s — about 44.64 GiB usable after driver and compositor overhead. 2050 of 2118 indexed models fit at 16K context with q4_0 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-8533
Bandwidth
273 GB/s
256-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 1761vision language 185image 2audio asr 39audio tts 21video 16embedding 26

What fits at 16K context

largest quantization that fits, per model · 2050 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
diffusiongemma-26B-A4B-it-HERETIC-UncensoredMoEBF1625.8B47.07 GiB0.26 GiB47.87 GiB0.13 GiB5±8.3%
diffusiongemma-26B-A4B-itMoEBF1625.8B47.07 GiB0.26 GiB47.87 GiB0.13 GiB5±8.3%
Devstral-2-123B-Instruct-2512UD-IQ3_XXS125B45.60 GiB1.55 GiB47.86 GiB0.14 GiB5±8.3%
gemma-4-26B-A4B-it-Claude-Opus-DistillMoEBF1626.5B47.04 GiB0.26 GiB47.83 GiB0.17 GiB5±8.3%
gemma-4-26B-A4B-it-Claude-Opus-Distill-v2MoEBF1626.5B47.04 GiB0.26 GiB47.83 GiB0.17 GiB5±8.3%
G4-MeroMero-26B-A4BMoEBF1625.8B47.04 GiB0.26 GiB47.83 GiB0.17 GiB5±8.3%
gemma-4-26B-A4B-it-qat-q4_0-unquantized-hereticMoEF1625.8B47.04 GiB0.26 GiB47.83 GiB0.17 GiB5±8.3%
G4-MeroMero-26B-A4B-it-uncensored-hereticMoEBF1625.8B47.04 GiB0.26 GiB47.83 GiB0.17 GiB5±8.3%
gemma-4-26B-A4B-itMoEF1626.5B47.04 GiB0.26 GiB47.83 GiB0.17 GiB5±8.3%
gemma-4-26B-A4B-it-ultra-uncensored-hereticMoEBF1625.8B47.04 GiB0.26 GiB47.83 GiB0.17 GiB5±8.3%
gemma-4-26B-A4B-it-uncensored-hereticMoEBF1625.8B47.04 GiB0.26 GiB47.83 GiB0.17 GiB5±8.3%
gemma-4-26B-A4B-it-abliterixMoEF1625.8B47.04 GiB0.26 GiB47.83 GiB0.17 GiB5±8.3%
gemma-4-26B-A4BMoEBF1626.5B47.04 GiB0.26 GiB47.83 GiB0.17 GiB5±8.3%
gemma-4-26B-A4B-Heretic-StableMoEBF1625.8B47.04 GiB0.26 GiB47.83 GiB0.17 GiB5±8.3%
gemma-4-26B-A4B-it-Uncensored-MAXMoEBF1625.8B47.04 GiB0.26 GiB47.83 GiB0.17 GiB5±8.3%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ3_K_S109B46.34 GiB0.84 GiB47.76 GiB0.24 GiB19±37%
Qwen3-Coder-NextMoEQ4_179.7B46.78 GiB0.42 GiB47.74 GiB0.26 GiB27±37%
Qwen3-Next-80B-A3B-ThinkingMoEQ4_181.3B46.78 GiB0.42 GiB47.74 GiB0.26 GiB27±37%
Qwen3-Next-80B-A3B-InstructMoEQ4_181.3B46.78 GiB0.42 GiB47.74 GiB0.26 GiB27±37%
Assistant_Pepe_70BQ5_K_S70.6B45.65 GiB1.41 GiB47.73 GiB0.27 GiB5±8.3%
Huihui-GLM-4.5-Air-abliterated-lossytensorsMoEI1-IQ3_XS110B46.34 GiB0.81 GiB47.72 GiB0.28 GiB19±37%
GLM-4.6VMoEUD-IQ3_XXS108B46.26 GiB0.81 GiB47.65 GiB0.35 GiB19±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEIQ2_XXS124B46.71 GiB0.39 GiB47.64 GiB0.36 GiB23±37%
HunyuanImage-2.1Q5_017.5B47.04 GiB0.00 GiB47.64 GiB0.36 GiB5±8.3%
Llama-3.1-70BQ5_070.6B45.45 GiB1.41 GiB47.53 GiB0.47 GiB5±8.3%
Mistral-Medium-3.5-128BIQ2_M128B45.27 GiB1.55 GiB47.52 GiB0.48 GiB5±8.3%
Apertus-70B-Instruct-2509Q5_K_S70.6B45.35 GiB1.41 GiB47.49 GiB0.51 GiB5±8.3%
CodeLlama-70b-Instruct-hfI1-Q5_K_M69.0B45.41 GiB1.41 GiB47.48 GiB0.52 GiB5±8.3%
CodeLlama-70b-Python-hfI1-Q5_K_M69.0B45.41 GiB1.41 GiB47.48 GiB0.52 GiB5±8.3%
Nous-Hermes-Llama2-70bI1-Q5_K_M69.0B45.41 GiB1.41 GiB47.48 GiB0.52 GiB5±8.3%
Midnight-Miqu-70B-v1.5I1-Q5_K_M69.0B45.41 GiB1.41 GiB47.48 GiB0.52 GiB5±8.3%
KafkaLM-70B-German-V0.1Q5_K_M69.0B45.41 GiB1.41 GiB47.48 GiB0.52 GiB5±8.3%
llama2_70b_chat_uncensoredQ5_K_M69.0B45.41 GiB1.41 GiB47.48 GiB0.52 GiB5±8.3%
Xwin-LM-70b-V0.1Q5_K_M69.0B45.41 GiB1.41 GiB47.48 GiB0.52 GiB5±8.3%
Llama-2-70b-chat-hfQ5_K_M69.0B45.41 GiB1.41 GiB47.48 GiB0.52 GiB5±8.3%
Meta-Llama-3-70B-InstructQ5_070.6B45.32 GiB1.41 GiB47.41 GiB0.59 GiB5±8.3%
Qwen3.5-122B-A10B-hereticMoEI1-IQ3_XS123B46.72 GiB0.11 GiB47.40 GiB0.60 GiB26±37%
Maenad-70BI1-Q5_K_S70.6B45.32 GiB1.41 GiB47.40 GiB0.60 GiB5±8.3%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-ReasonerI1-Q5_K_S70.6B45.32 GiB1.41 GiB47.40 GiB0.60 GiB5±8.3%
calme-2.4-llama3-70bQ5_K_S70.6B45.32 GiB1.41 GiB47.40 GiB0.60 GiB5±8.3%
calme-2.2-llama3-70bQ5_K_S70.6B45.32 GiB1.41 GiB47.40 GiB0.60 GiB5±8.3%
Rombos-LLM-70b-Llama-3.3I1-Q5_K_S70.6B45.32 GiB1.41 GiB47.40 GiB0.60 GiB5±8.3%
L3.3-Electra-R1-70bI1-Q5_K_S70.6B45.32 GiB1.41 GiB47.40 GiB0.60 GiB5±8.3%
L3.3-70B-Magnum-v4-SEQ5_K_S70.6B45.32 GiB1.41 GiB47.40 GiB0.60 GiB5±8.3%
Latxa-Llama-3.1-70B-Instruct-v2I1-Q5_K_S70.6B45.32 GiB1.41 GiB47.40 GiB0.60 GiB5±8.3%
Llama-3.3_70_b_uncensored_continuedI1-Q5_K_S70.6B45.32 GiB1.41 GiB47.40 GiB0.60 GiB5±8.3%
Llama-3.3-70B-Instruct-abliteratedI1-Q5_K_S70.6B45.32 GiB1.41 GiB47.40 GiB0.60 GiB5±8.3%
Strawberrylemonade-L3-70B-v1.2Q5_K_S70.6B45.32 GiB1.41 GiB47.40 GiB0.60 GiB5±8.3%
grok-oss-Revenant-70BI1-Q5_K_S70.6B45.32 GiB1.41 GiB47.40 GiB0.60 GiB5±8.3%
Llama-3.1-Nemotron-70B-Instruct-HFI1-Q5_K_S70.6B45.32 GiB1.41 GiB47.40 GiB0.60 GiB5±8.3%
L3.3-70B-Euryale-v2.3I1-Q5_K_S70.6B45.32 GiB1.41 GiB47.40 GiB0.60 GiB5±8.3%
Hermes-4-70B-hereticI1-Q5_K_S70.6B45.32 GiB1.41 GiB47.40 GiB0.60 GiB5±8.3%
Hermes-4-70BQ5_K_S70.6B45.32 GiB1.41 GiB47.40 GiB0.60 GiB5±8.3%
Llama-3.3-70B-InstructQ5_K_S70.6B45.32 GiB1.41 GiB47.40 GiB0.60 GiB5±8.3%
Hermes-3-Llama-3.1-70BQ5_K_S70.6B45.32 GiB1.41 GiB47.40 GiB0.60 GiB5±8.3%
Anubis-70B-v1.2Q5_K_S70.6B45.32 GiB1.41 GiB47.40 GiB0.60 GiB5±8.3%
Golem-70B-v1bI1-Q5_K_S70.6B45.32 GiB1.41 GiB47.40 GiB0.60 GiB5±8.3%
DeepSeek-R1-Distill-Llama-70B-abliteratedI1-Q5_K_S70.6B45.32 GiB1.41 GiB47.40 GiB0.60 GiB5±8.3%
DeepSeek-R1-Distill-Llama-70B-hereticI1-Q5_K_S70.6B45.32 GiB1.41 GiB47.40 GiB0.60 GiB5±8.3%
DeepSeek-R1-Distill-Llama-70BQ5_K_S70.6B45.32 GiB1.41 GiB47.40 GiB0.60 GiB5±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 Pro run?
2050 of 2118 indexed open-weight models fit a Apple M4 Pro at 16,384 context with q4_0 KV cache, the largest being diffusiongemma-26B-A4B-it-HERETIC-Uncensored at BF16. That covers text, vision-language, image, video and speech models.
How much usable memory does a Apple M4 Pro 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 M4 Pro fast for local AI?
Its memory bandwidth is 273 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.