Apple · apple

Apple M4 Pro

Apple M4 Pro has 48 GB of unified memory at 273 GB/s — about 33.48 GiB usable after driver and compositor overhead. 2030 of 2118 indexed models fit at 4K 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
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 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◐
CodeLlama-70b-Python-hfIQ4_XS69.0B34.64 GiB0.66 GiB35.98 GiB0.02 GiB6±8.3%
Midnight-Miqu-70B-v1.5IQ4_XS69.0B34.64 GiB0.66 GiB35.98 GiB0.02 GiB6±8.3%
Qwen3-53B-A3B-2507-THINKING-TOTAL-RECALL-v2-MASTER-CODERMoEI1-Q5_K_M53.0B35.08 GiB0.35 GiB35.97 GiB0.03 GiB23±37%
Meta-Llama-3-70B-InstructQ3_K_L70.6B34.60 GiB0.66 GiB35.94 GiB0.06 GiB6±8.3%
Maenad-70BI1-Q3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-ReasonerI1-Q3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
calme-2.4-llama3-70bQ3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
calme-2.2-llama3-70bQ3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
Rombos-LLM-70b-Llama-3.3I1-Q3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
L3.3-Electra-R1-70bI1-Q3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
L3.3-70B-Magnum-v4-SEQ3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
Latxa-Llama-3.1-70B-Instruct-v2I1-Q3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
Llama-3.3_70_b_uncensored_continuedI1-Q3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
Llama-3.3-70B-Instruct-abliteratedI1-Q3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
Strawberrylemonade-L3-70B-v1.2Q3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
grok-oss-Revenant-70BI1-Q3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
Llama-3.1-Nemotron-70B-Instruct-HFI1-Q3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
L3.3-70B-Euryale-v2.3I1-Q3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
Hermes-4-70B-hereticI1-Q3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
Llama-3.3-70B-InstructQ3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
Llama-3.1-70BQ3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
Hermes-3-Llama-3.1-70BQ3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
Anubis-70B-v1.2Q3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
Hermes-4-70BQ3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
Golem-70B-v1bI1-Q3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
DeepSeek-R1-Distill-Llama-70B-abliteratedI1-Q3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
DeepSeek-R1-Distill-Llama-70B-hereticI1-Q3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
llama-3-firefunction-v2Q3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
DeepSeek-R1-Distill-Llama-70BQ3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
Legion-V2.1-LLaMa-70BI1-Q3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
Assistant_Pepe_70BI1-Q3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
Tess-R1-Limerick-Llama-3.1-70BQ3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
SEMIKONG-70BQ3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
functionary-medium-v3.2KV unresolvedQ3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
Llama-3.1-WhiteRabbitNeo-2-70BQ3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
Infinity-Instruct-7M-Gen-Llama3_1-70BI1-Q3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
New-Dawn-Llama-3-70B-32K-v1.0I1-Q3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
Meta-Llama-3-70B-Instruct-abliterated-v3.5I1-Q3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
Athene-70BQ3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
L3.3-70B-Magnum-DiamondQ3_K_L70.6B34.59 GiB0.66 GiB35.93 GiB0.07 GiB6±8.3%
Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16MoEIQ4_XS35.1B35.30 GiB0.04 GiB35.90 GiB0.10 GiB33±37%
Qwen2.5-Coder-32B-InstructQ4_032.8B34.72 GiB0.53 GiB35.90 GiB0.10 GiB6±8.3%
Qwen3-Coder-Next-REAMMoEI1-Q4_160.3B35.30 GiB0.05 GiB35.89 GiB0.11 GiB35±37%
Salience-1.5-ProMoEQ8_036.0B35.22 GiB0.04 GiB35.81 GiB0.19 GiB33±37%
Qwable-v1MoEQ8_036.0B35.22 GiB0.04 GiB35.81 GiB0.19 GiB33±37%
T-SearchMoEQ8_036.0B35.22 GiB0.04 GiB35.81 GiB0.19 GiB33±37%
Qwen3.5-35B-A3BMoEQ8_036.0B35.22 GiB0.04 GiB35.81 GiB0.19 GiB33±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedUD-IQ2_XXS109B34.83 GiB0.40 GiB35.81 GiB0.19 GiB25±37%
Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-PreservedMoEQ8_035.1B35.21 GiB0.04 GiB35.80 GiB0.20 GiB33±37%
Qwen3.6-35B-A3B-Fable-5-DistillMoEQ8_036.0B35.21 GiB0.04 GiB35.80 GiB0.20 GiB33±37%
Qwable-v2MoEQ8_036.0B35.21 GiB0.04 GiB35.80 GiB0.20 GiB33±37%
Qwen3.6-35B-A3B-YOYO-V2MoEQ8_036.0B35.21 GiB0.04 GiB35.80 GiB0.20 GiB33±37%
Ornith-1.0-35B-FP8-BLOCK-MTPMoEQ8_035.5B35.21 GiB0.04 GiB35.80 GiB0.20 GiB33±37%
fable-coder-35B-A3BMoEQ8_036.0B35.21 GiB0.04 GiB35.80 GiB0.20 GiB33±37%
PINQWEN-3.6-35B-CLEAN-BF16MoEQ8_036.0B35.21 GiB0.04 GiB35.80 GiB0.20 GiB33±37%
UniMath-35B-A3BMoEQ8_036.0B35.21 GiB0.04 GiB35.80 GiB0.20 GiB33±37%
Ornith-1.0-35B-Heretic-MTPMoEQ8_0—35.21 GiB0.04 GiB35.80 GiB0.20 GiB33±37%
Fawen-1.0-35BMoEQ8_036.0B35.21 GiB0.04 GiB35.80 GiB0.20 GiB33±37%
Qwopus3.6-35B-A3B-v1MoEQ8_036.0B35.21 GiB0.04 GiB35.80 GiB0.20 GiB33±37%
CyberStrike-OffSec-35BMoEQ8_035.1B35.21 GiB0.04 GiB35.80 GiB0.20 GiB33±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 Pro run?
2030 of 2118 indexed open-weight models fit a Apple M4 Pro at 4,096 context with q8_0 KV cache, the largest being CodeLlama-70b-Python-hf at IQ4_XS. That covers text, vision-language, image, video and speech models.
How much usable memory does a Apple M4 Pro 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 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.