Apple · apple

Apple M4

Apple M4 has 12 GB of unified memory at 120 GB/s — about 8.37 GiB usable after driver and compositor overhead. 1603 of 2118 indexed models fit at 4K context with f16 KV. Note only 9 GB of its 12 GB is allocatable to the GPU.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
12 GB
LPDDR5X-7500
Bandwidth
120 GB/s
128-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
embedding 26text 1386video 12vision language 117audio asr 39audio tts 21image 2

What fits at 4K context

largest quantization that fits, per model · 1603 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Nemotron-3-Embed-8B-BF16Q8_08.0B7.88 GiB0.53 GiB9.00 GiB0.00 GiB11±8.3%
DeepSeek-Coder-V2-Lite-BaseMoEI1-Q4_015.7B8.32 GiB0.12 GiB9.00 GiB0.00 GiB34±37%
Wan2.1-T2V-14BQ4_014.3B8.41 GiB0.00 GiB9.00 GiB0.00 GiB11±8.3%
codellama-13b-oasst-sft-v10Q3_K_S13.0B5.27 GiB3.13 GiB8.99 GiB0.01 GiB11±8.3%
chronos-hermes-13b-v2Q3_K_S13.0B5.27 GiB3.13 GiB8.99 GiB0.01 GiB11±8.3%
WhiteRabbitNeo-13B-v1Q3_K_S13.0B5.27 GiB3.13 GiB8.99 GiB0.01 GiB11±8.3%
CodeLlama-13b-Instruct-hfQ3_K_S13.0B5.27 GiB3.13 GiB8.99 GiB0.01 GiB11±8.3%
Orca-2-13b-Alpaca-UncensoredI1-IQ3_S13.0B5.27 GiB3.13 GiB8.99 GiB0.01 GiB11±8.3%
WizardLM-13B-UncensoredI1-IQ3_S13.0B5.27 GiB3.13 GiB8.99 GiB0.01 GiB11±8.3%
WizardCoder-Python-13B-V1.0I1-IQ3_S13.0B5.27 GiB3.13 GiB8.99 GiB0.01 GiB11±8.3%
Guanaco-13B-UncensoredI1-IQ3_S13.0B5.27 GiB3.13 GiB8.99 GiB0.01 GiB11±8.3%
Llama-2-13b-chat-hfQ3_K_S13.0B5.27 GiB3.13 GiB8.99 GiB0.01 GiB11±8.3%
Wizard-Vicuna-13B-UncensoredQ3_K_S13.0B5.27 GiB3.13 GiB8.99 GiB0.01 GiB11±8.3%
WizardLM-13b-V1.0-UncensoredQ3_K_S13.0B5.27 GiB3.13 GiB8.99 GiB0.01 GiB11±8.3%
MythoMax-L2-13bQ3_K_S13.0B5.27 GiB3.13 GiB8.99 GiB0.01 GiB11±8.3%
WizardLM-1.0-Uncensored-Llama2-13bQ3_K_S13.0B5.27 GiB3.13 GiB8.99 GiB0.01 GiB11±8.3%
speechless-llama2-hermes-orca-platypus-wizardlm-13bQ3_K_S13.0B5.27 GiB3.13 GiB8.99 GiB0.01 GiB11±8.3%
mythalion-13bQ3_K_S13.0B5.27 GiB3.13 GiB8.99 GiB0.01 GiB11±8.3%
UncensoredLM-DeepSeek-R1-Distill-Qwen-14BIQ4_NL14.2B7.67 GiB0.72 GiB8.99 GiB0.01 GiB11±8.3%
Qwen3-VL-30B-A3B-ThinkingMoEIQ2_XS31.1B8.07 GiB0.38 GiB8.99 GiB0.01 GiB35±37%
MiroThinker-v1.0-30BMoEIQ2_XS30.5B8.07 GiB0.38 GiB8.99 GiB0.01 GiB35±37%
Qwen3-30B-A3B-Instruct-2507MoEIQ2_XS30.5B8.07 GiB0.38 GiB8.99 GiB0.01 GiB35±37%
Qwen3-30B-A3B-Thinking-2507MoEIQ2_XS30.5B8.07 GiB0.38 GiB8.99 GiB0.01 GiB35±37%
Llama3.2-30B-A3B-II-Dark-Champion-INSTRUCT-Heretic-Abliterated-UncensoredMoEI1-IQ2_XXS30.0B7.69 GiB0.73 GiB8.99 GiB0.01 GiB24±37%
Tini-Cybersec-8B-A1BMoEQ8_08.5B8.39 GiB0.05 GiB8.99 GiB0.01 GiB31±37%
LFM2.5-8B-A1B-KO-SFTMoEQ8_08.5B8.39 GiB0.05 GiB8.99 GiB0.01 GiB31±37%
LFM2.5-8B-A1B-SOMPOA-heresyMoEQ8_08.5B8.39 GiB0.05 GiB8.99 GiB0.01 GiB31±37%
Huihui-LFM2.5-8B-A1B-abliteratedMoEQ8_08.5B8.39 GiB0.05 GiB8.99 GiB0.01 GiB31±37%
LFM2.5-8B-A1BMoEQ8_08.5B8.39 GiB0.05 GiB8.99 GiB0.01 GiB31±37%
Supertron2.1-8B-A1BMoEQ8_08.5B8.39 GiB0.05 GiB8.99 GiB0.01 GiB31±37%
LFM2.5-8B-A1B-hereticMoEQ8_08.5B8.39 GiB0.05 GiB8.99 GiB0.01 GiB31±37%
Tongyi-DeepResearch-30B-A3BMoEIQ2_XS30.5B8.07 GiB0.38 GiB8.98 GiB0.02 GiB35±37%
gemma-3-12b-it-vl-Gemini-3-Pro-Preview-Heretic-Uncensored-ThinkingI1-Q5_K_S12.2B7.67 GiB0.72 GiB8.98 GiB0.02 GiB11±8.3%
gemma-3-12b-it-vl-Deepseek-v3.1-Heretic-Uncensored-ThinkingI1-Q5_K_S12.2B7.67 GiB0.72 GiB8.98 GiB0.02 GiB11±8.3%
gemma-3-12b-it-ultra-uncensored-hereticQ5_K_S12.2B7.67 GiB0.72 GiB8.98 GiB0.02 GiB11±8.3%
gemma-3-12b-it-vl-GLM-4.7-Flash-Heretic-Uncensored-ThinkingI1-Q5_K_S12.2B7.67 GiB0.72 GiB8.98 GiB0.02 GiB11±8.3%
Floppa-12B-Gemma3-UncensoredI1-Q5_K_S12.2B7.67 GiB0.72 GiB8.98 GiB0.02 GiB11±8.3%
gemma-3-12b-it-hereticI1-Q5_K_S12.2B7.67 GiB0.72 GiB8.98 GiB0.02 GiB11±8.3%
gemma-3-12b-it-abliteratedQ5_K_S12.2B7.67 GiB0.72 GiB8.98 GiB0.02 GiB11±8.3%
gemma-3-12b-it-abliterated-v2Q5_K_S11.8B7.67 GiB0.72 GiB8.98 GiB0.02 GiB11±8.3%
gemma-3-12b-itQ5_K_S12.2B7.67 GiB0.72 GiB8.98 GiB0.02 GiB11±8.3%
InternVL3_5-14BQ4_K_M15.1B8.38 GiB0.00 GiB8.98 GiB0.02 GiB11±8.3%
Magistry-24B-v1.1IQ2_S23.6B7.68 GiB0.63 GiB8.97 GiB0.03 GiB12±8.3%
diffusiongemma-26B-A4B-it-HERETIC-UncensoredMoETQ1_025.8B7.99 GiB0.45 GiB8.97 GiB0.03 GiB11±8.3%
gemma-4-12B-it-uncensored-hereticQ4_K_S12.0B7.66 GiB0.72 GiB8.97 GiB0.03 GiB12±8.3%
DeepSeek-Coder-V2-Lite-InstructMoEIQ4_NL15.7B8.29 GiB0.12 GiB8.97 GiB0.03 GiB35±37%
DeepSeek-V2-Lite-ChatMoEIQ4_NL15.7B8.29 GiB0.12 GiB8.97 GiB0.03 GiB35±37%
Olmo-3.1-32B-InstructUD-IQ1_M32.2B7.33 GiB1.00 GiB8.97 GiB0.03 GiB12±8.3%
Olmo-3.1-32B-ThinkUD-IQ1_M32.2B7.33 GiB1.00 GiB8.97 GiB0.03 GiB12±8.3%
Olmo-3-32B-ThinkUD-IQ1_M32.2B7.33 GiB1.00 GiB8.97 GiB0.03 GiB12±8.3%
HunyuanVideo-1.5Q8_08.3B8.38 GiB0.00 GiB8.97 GiB0.03 GiB12±8.3%
Devstral-Small-2-24B-Instruct-2512UD-IQ2_M24.0B7.68 GiB0.63 GiB8.97 GiB0.03 GiB12±8.3%
Qwen2.5-14B-Instruct-abliterated-v2IQ4_XS14.8B7.62 GiB0.75 GiB8.97 GiB0.03 GiB12±8.3%
Qwen2.5-14B-Instruct-UncensoredIQ4_XS14.8B7.62 GiB0.75 GiB8.97 GiB0.03 GiB12±8.3%
Qwen2.5-14B-Instruct-1M-abliteratedIQ4_XS14.8B7.62 GiB0.75 GiB8.97 GiB0.03 GiB12±8.3%
Ektome-Qwen2.5-Coder-14B-Instruct-PristinelyUncensoredIQ4_XS14.8B7.62 GiB0.75 GiB8.97 GiB0.03 GiB12±8.3%
14B-Qwen2.5-Kunou-v1IQ4_XS14.8B7.62 GiB0.75 GiB8.97 GiB0.03 GiB12±8.3%
FinetunedQwen14BIQ4_XS14.8B7.62 GiB0.75 GiB8.97 GiB0.03 GiB12±8.3%
DeepSeek-R1-Distill-Qwen-14B-abliterated-v2IQ4_XS14.8B7.62 GiB0.75 GiB8.97 GiB0.03 GiB12±8.3%
C1-TachuIQ4_XS14.8B7.62 GiB0.75 GiB8.97 GiB0.03 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 run?
1603 of 2118 indexed open-weight models fit a Apple M4 at 4,096 context with f16 KV cache, the largest being Nemotron-3-Embed-8B-BF16 at Q8_0. That covers text, vision-language, image, video and speech models.
How much usable memory does a Apple M4 actually have?
Its nameplate is 12 GB, but about 8.37 GiB is available to a model once driver and compositor overhead is accounted for, and only 9 GB of the pool can be allocated to the GPU at all.
Is a Apple M4 fast for local AI?
Its memory bandwidth is 120 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.