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. 2029 of 2118 indexed models fit at 32K 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
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
vision language 183text 1742video 16audio tts 21image 2embedding 26audio asr 39

What fits at 32K context

largest quantization that fits, per model · 2029 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-PreservedQ8_027.4B34.80 GiB0.56 GiB35.97 GiB0.03 GiB6±8.3%
Salience-1.5-ProMoEQ8_036.0B35.22 GiB0.18 GiB35.95 GiB0.05 GiB32±37%
Qwable-v1MoEQ8_036.0B35.22 GiB0.18 GiB35.95 GiB0.05 GiB32±37%
T-SearchMoEQ8_036.0B35.22 GiB0.18 GiB35.95 GiB0.05 GiB32±37%
Qwen3.5-35B-A3BMoEQ8_036.0B35.22 GiB0.18 GiB35.95 GiB0.05 GiB32±37%
Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-PreservedMoEQ8_035.1B35.21 GiB0.18 GiB35.94 GiB0.06 GiB32±37%
Qwen3.6-35B-A3B-Fable-5-DistillMoEQ8_036.0B35.21 GiB0.18 GiB35.94 GiB0.06 GiB32±37%
Qwable-v2MoEQ8_036.0B35.21 GiB0.18 GiB35.94 GiB0.06 GiB32±37%
Qwen3.6-35B-A3B-YOYO-V2MoEQ8_036.0B35.21 GiB0.18 GiB35.94 GiB0.06 GiB32±37%
Ornith-1.0-35B-FP8-BLOCK-MTPMoEQ8_035.5B35.21 GiB0.18 GiB35.94 GiB0.06 GiB32±37%
fable-coder-35B-A3BMoEQ8_036.0B35.21 GiB0.18 GiB35.94 GiB0.06 GiB32±37%
PINQWEN-3.6-35B-CLEAN-BF16MoEQ8_036.0B35.21 GiB0.18 GiB35.94 GiB0.06 GiB32±37%
UniMath-35B-A3BMoEQ8_036.0B35.21 GiB0.18 GiB35.94 GiB0.06 GiB32±37%
Ornith-1.0-35B-Heretic-MTPMoEQ8_0—35.21 GiB0.18 GiB35.94 GiB0.06 GiB32±37%
Fawen-1.0-35BMoEQ8_036.0B35.21 GiB0.18 GiB35.94 GiB0.06 GiB32±37%
Qwopus3.6-35B-A3B-v1MoEQ8_036.0B35.21 GiB0.18 GiB35.94 GiB0.06 GiB32±37%
CyberStrike-OffSec-35BMoEQ8_035.1B35.21 GiB0.18 GiB35.94 GiB0.06 GiB32±37%
Qwen3.6-35B-A3BMoEQ8_036.0B35.21 GiB0.18 GiB35.94 GiB0.06 GiB32±37%
Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-DistilledMoEQ8_036.0B35.21 GiB0.18 GiB35.94 GiB0.06 GiB32±37%
Qwen3.5-35B-A3B-uncensored-heretic-v2-Native-MTP-PreservedMoEQ8_035.1B35.21 GiB0.18 GiB35.94 GiB0.06 GiB32±37%
llama-3.2-3b-instructF163.2B34.37 GiB0.98 GiB35.92 GiB0.08 GiB6±8.3%
Llama-3_3-Nemotron-Super-49B-v1_5IQ2_XXS49.9B12.72 GiB22.50 GiB35.91 GiB0.09 GiB6±8.3%
Valkyrie-49B-v2.1I1-IQ2_XXS49.9B12.72 GiB22.50 GiB35.91 GiB0.09 GiB6±8.3%
Llama-3_3-Nemotron-Super-49B-v1IQ2_XXS49.9B12.72 GiB22.50 GiB35.91 GiB0.09 GiB6±8.3%
Devstral-2-123B-Instruct-2512UD-IQ2_XXS125B32.04 GiB3.09 GiB35.84 GiB0.16 GiB6±8.3%
deepseek-coder-33b-instructQ8_033.3B33.00 GiB2.18 GiB35.81 GiB0.19 GiB6±8.3%
deepseek-coder-33b-baseQ8_033.3B33.00 GiB2.18 GiB35.81 GiB0.19 GiB6±8.3%
WhiteRabbitNeo-33B-v1Q8_033.3B33.00 GiB2.18 GiB35.81 GiB0.19 GiB6±8.3%
v6-Finch-14B-HFF1614.1B26.63 GiB8.58 GiB35.81 GiB0.19 GiB6±8.3%
WizardLM-Uncensored-SuperCOT-StoryTelling-30bQ5_K_M32.5B21.46 GiB13.71 GiB35.80 GiB0.20 GiB6±8.3%
Wizard-Vicuna-30B-UncensoredI1-Q5_K_M32.5B21.46 GiB13.71 GiB35.80 GiB0.20 GiB6±8.3%
archangel_sft-kto_llama30bI1-Q5_K_M32.5B21.46 GiB13.71 GiB35.80 GiB0.20 GiB6±8.3%
Mistral-Small-4-119B-2603MoEUD-IQ2_M119B34.99 GiB0.20 GiB35.77 GiB0.23 GiB32±37%
Laguna-S-2.1MoEUD-IQ2_M118B34.71 GiB0.46 GiB35.74 GiB0.26 GiB29±37%
WizardCoder-Python-34B-V1.0Q8_033.7B33.39 GiB1.69 GiB35.73 GiB0.27 GiB6±8.3%
Phind-CodeLlama-34B-v2Q8_033.7B33.39 GiB1.69 GiB35.73 GiB0.27 GiB6±8.3%
CodeLlama-34b-instruct-hfQ8_033.7B33.39 GiB1.69 GiB35.73 GiB0.27 GiB6±8.3%
WizardLM-1.0-Uncensored-CodeLlama-34bQ8_033.7B33.39 GiB1.69 GiB35.73 GiB0.27 GiB6±8.3%
Phind-CodeLlama-34B-Python-v1Q8_033.7B33.39 GiB1.69 GiB35.73 GiB0.27 GiB6±8.3%
GLM-4.6VMoEIQ1_M108B33.46 GiB1.62 GiB35.65 GiB0.35 GiB21±37%
Rombo-LLM-V3.0-Qwen-72bI1-IQ3_S72.7B32.12 GiB2.81 GiB35.61 GiB0.39 GiB6±8.3%
Qwen2.5-72B-Instruct-abliteratedI1-IQ3_S72.7B32.12 GiB2.81 GiB35.61 GiB0.39 GiB6±8.3%
Qwen2.5-72B-Instruct-abliterated-v2I1-IQ3_S72.7B32.12 GiB2.81 GiB35.61 GiB0.39 GiB6±8.3%
HuatuoGPT-o1-72BQ3_K_S72.7B32.12 GiB2.81 GiB35.61 GiB0.39 GiB6±8.3%
MiroThinker-v1.0-72BI1-IQ3_S72.7B32.12 GiB2.81 GiB35.61 GiB0.39 GiB6±8.3%
EVA-Qwen2.5-72B-v0.2Q3_K_S72.7B32.12 GiB2.81 GiB35.61 GiB0.39 GiB6±8.3%
Qwen2.5-Math-72B-InstructQ3_K_S72.7B32.12 GiB2.81 GiB35.61 GiB0.39 GiB6±8.3%
Qwen2.5-72B-InstructQ3_K_S72.7B32.12 GiB2.81 GiB35.61 GiB0.39 GiB6±8.3%
Malaysian-Qwen2.5-72B-InstructI1-IQ3_S72.7B32.12 GiB2.81 GiB35.61 GiB0.39 GiB6±8.3%
Qwen2.5-72BI1-IQ3_S72.7B32.12 GiB2.81 GiB35.61 GiB0.39 GiB6±8.3%
magnum-v4-72bI1-IQ3_S72.7B32.12 GiB2.81 GiB35.61 GiB0.39 GiB6±8.3%
Kimi-Dev-72BQ3_K_S72.7B32.12 GiB2.81 GiB35.61 GiB0.39 GiB6±8.3%
KAT-Dev-72B-ExpQ3_K_S72.7B32.12 GiB2.81 GiB35.61 GiB0.39 GiB6±8.3%
Homer-v1.0-Qwen2.5-72BQ3_K_S72.7B32.12 GiB2.81 GiB35.61 GiB0.39 GiB6±8.3%
Chuluun-Qwen2.5-72B-v0.01Q3_K_S72.7B32.12 GiB2.81 GiB35.61 GiB0.39 GiB6±8.3%
Qwen2.5-VL-72B-InstructQ3_K_S73.4B32.12 GiB2.81 GiB35.61 GiB0.39 GiB6±8.3%
Tower-Plus-72B-ultra-uncensored-hereticI1-IQ3_S72.7B32.12 GiB2.81 GiB35.61 GiB0.39 GiB6±8.3%
Chronos-Platinum-72BQ3_K_S72.7B32.12 GiB2.81 GiB35.61 GiB0.39 GiB6±8.3%
UI-TARS-72B-DPOQ3_K_S73.4B32.12 GiB2.81 GiB35.61 GiB0.39 GiB6±8.3%
GLM-Z1-Rumination-32B-0414Q8_033.1B32.81 GiB2.14 GiB35.59 GiB0.41 GiB6±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?
2029 of 2118 indexed open-weight models fit a Apple M4 Pro at 32,768 context with q4_0 KV cache, the largest being Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved at Q8_0. 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.