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. 1542 of 2118 indexed models fit at 16K context with q8_0 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
text 1328vision language 114video 12audio asr 39image 2audio tts 21embedding 26

What fits at 16K context

largest quantization that fits, per model · 1542 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Qwen3.6-27B-Heretic2-ThinkingI1-IQ2_XXS27.4B7.85 GiB0.53 GiB9.00 GiB0.00 GiB11±8.3%
Qwen3.6-27B-Uncensored-AggressiveI1-IQ2_XXS27.4B7.85 GiB0.53 GiB9.00 GiB0.00 GiB11±8.3%
Qwen-3.5-Opus-GLM-27BI1-IQ2_XXS26.9B7.85 GiB0.53 GiB9.00 GiB0.00 GiB11±8.3%
Qwen3.6-27B-abliteratedI1-IQ2_XXS27.4B7.85 GiB0.53 GiB9.00 GiB0.00 GiB11±8.3%
KoQweopus-3.5-27B-experimentalI1-IQ2_XXS27.8B7.85 GiB0.53 GiB9.00 GiB0.00 GiB11±8.3%
Webcoda-AI-27BI1-IQ2_XXS27.4B7.85 GiB0.53 GiB9.00 GiB0.00 GiB11±8.3%
Qwen3.5-27B-imabari-v2I1-IQ2_XXS27.8B7.85 GiB0.53 GiB9.00 GiB0.00 GiB11±8.3%
Qwen3.5-27B-uncensored-heretic-v1I1-IQ2_XXS27.4B7.85 GiB0.53 GiB9.00 GiB0.00 GiB11±8.3%
Carnice-V2-27bI1-IQ2_XXS27.4B7.85 GiB0.53 GiB9.00 GiB0.00 GiB11±8.3%
Qwen3.5-Queen-27BI1-IQ2_XXS27.4B7.85 GiB0.53 GiB9.00 GiB0.00 GiB11±8.3%
GRaPE-2-ProI1-IQ2_XXS27.8B7.85 GiB0.53 GiB9.00 GiB0.00 GiB11±8.3%
Qwen3.6-27B-Omnimerge-v4IQ2_XXS27.8B7.85 GiB0.53 GiB9.00 GiB0.00 GiB11±8.3%
Qwen3.6-27B-AblitIQ2_XXS26.9B7.85 GiB0.53 GiB9.00 GiB0.00 GiB11±8.3%
Darwin-28B-REASONI1-IQ2_XXS26.9B7.85 GiB0.53 GiB9.00 GiB0.00 GiB11±8.3%
Huihui-Qwen3.5-27B-Claude-4.6-Opus-abliteratedI1-IQ2_XXS27.8B7.85 GiB0.53 GiB9.00 GiB0.00 GiB11±8.3%
Qwen3.5-27B-WebNovel-Writer-zhI1-IQ2_XXS26.9B7.85 GiB0.53 GiB9.00 GiB0.00 GiB11±8.3%
Qwen3.5-27B_Homebrew-v2I1-IQ2_XXS27.4B7.85 GiB0.53 GiB9.00 GiB0.00 GiB11±8.3%
Wan2.1-T2V-14BQ4_014.3B8.41 GiB0.00 GiB9.00 GiB0.00 GiB11±8.3%
WizardCoder-Python-34B-V1.0I1-IQ1_S33.7B6.75 GiB1.59 GiB8.99 GiB0.01 GiB12±8.3%
Phind-CodeLlama-34B-Python-v1I1-IQ1_S33.7B6.75 GiB1.59 GiB8.99 GiB0.01 GiB12±8.3%
Phind-CodeLlama-34B-v2I1-IQ1_S33.7B6.75 GiB1.59 GiB8.99 GiB0.01 GiB12±8.3%
OLMo-2-1124-7B-InstructQ4_K_M7.3B4.16 GiB4.25 GiB8.99 GiB0.01 GiB11±8.3%
gemma-3n-E2B-itF165.4B8.31 GiB0.12 GiB8.99 GiB0.01 GiB11±8.3%
gemma-4-12B-it-qat-q4_0-unquantized-uncensored-hereticNVFP412.0B7.60 GiB0.78 GiB8.98 GiB0.02 GiB11±8.3%
gemma-4-12B-it-uncensored-hereticNVFP412.0B7.60 GiB0.78 GiB8.98 GiB0.02 GiB11±8.3%
InternVL3_5-14BQ4_K_M15.1B8.38 GiB0.00 GiB8.98 GiB0.02 GiB11±8.3%
AMALIA-9B-0626-DPOQ6_K9.2B7.00 GiB1.39 GiB8.97 GiB0.03 GiB11±8.3%
HunyuanVideo-1.5Q8_08.3B8.38 GiB0.00 GiB8.97 GiB0.03 GiB12±8.3%
GigaChat3-10B-A1.8B-baseMoEQ6_K11.5B8.18 GiB0.24 GiB8.97 GiB0.03 GiB37±37%
Goetia-26B-A4B-v1.4MoEI1-IQ1_S26.0B7.95 GiB0.49 GiB8.97 GiB0.03 GiB11±8.3%
G4-Moonlight-Dusk-26B-A4B-hereticMoEI1-IQ1_S26.5B7.95 GiB0.49 GiB8.97 GiB0.03 GiB11±8.3%
Pantheon-Reasoning-26B-A4B-1.1-hereticMoEI1-IQ1_S26.5B7.95 GiB0.49 GiB8.97 GiB0.03 GiB11±8.3%
G4-Moonlight-Dusk-26B-A4BMoEI1-IQ1_S26.5B7.95 GiB0.49 GiB8.97 GiB0.03 GiB11±8.3%
Chimera-X-26B-A4BMoEI1-IQ1_S26.5B7.95 GiB0.49 GiB8.97 GiB0.03 GiB11±8.3%
Pantheon-Reasoning-26B-A4B-1.1MoEI1-IQ1_S26.5B7.95 GiB0.49 GiB8.97 GiB0.03 GiB11±8.3%
Gemma-4-26B-A4B-StyleTune-V2MoEI1-IQ1_S26.5B7.95 GiB0.49 GiB8.97 GiB0.03 GiB11±8.3%
Gemma-4-26B-A4B-StyleTuneMoEI1-IQ1_S26.5B7.95 GiB0.49 GiB8.97 GiB0.03 GiB11±8.3%
gemma-4-26b-a4b-heretic-styletune-v2-headMoEI1-IQ1_S25.8B7.95 GiB0.49 GiB8.97 GiB0.03 GiB11±8.3%
Qwen3.5-27B-Engineer-Deckard-GeminiI1-IQ2_XS27.7B7.83 GiB0.53 GiB8.97 GiB0.03 GiB12±8.3%
Qwen3.5-27B-HERETIC-Polaris-Advanced-Thinking-Alpha-uncensoredI1-IQ2_XS27.4B7.83 GiB0.53 GiB8.97 GiB0.03 GiB12±8.3%
Qwen3.5-27B-Deckard-PKD-Heretic-Uncensored-ThinkingI1-IQ2_XS27.4B7.83 GiB0.53 GiB8.97 GiB0.03 GiB12±8.3%
Huihui-Qwen3.5-27B-abliteratedI1-IQ2_XS27.8B7.83 GiB0.53 GiB8.97 GiB0.03 GiB12±8.3%
Qwen3.5-27B-Unredacted-MAXI1-IQ2_XS27.4B7.83 GiB0.53 GiB8.97 GiB0.03 GiB12±8.3%
Qwen3.5-27B-hereticI1-IQ2_XS27.4B7.83 GiB0.53 GiB8.97 GiB0.03 GiB12±8.3%
Qwen3.5-27B-DerestrictedI1-IQ2_XS27.8B7.83 GiB0.53 GiB8.97 GiB0.03 GiB12±8.3%
Qwen3.5-27B-Claude-4.6-Opus-Reasoning-DistilledI1-IQ2_XS27.8B7.83 GiB0.53 GiB8.97 GiB0.03 GiB12±8.3%
GLM-4.7-Flash-REAP-23B-A3BMoEUD-IQ2_M23.0B7.97 GiB0.44 GiB8.97 GiB0.03 GiB31±37%
NVIDIA-Nemotron-Nano-12B-v2IQ4_XS12.3B6.29 GiB2.06 GiB8.97 GiB0.03 GiB12±8.3%
Nemotron-Mini-4B-InstructQ4_04.2B7.34 GiB1.06 GiB8.96 GiB0.04 GiB11±8.3%
Voxtral-Small-24B-2507IQ2_S24.3B6.97 GiB1.33 GiB8.96 GiB0.04 GiB12±8.3%
Devstral-Small-2-24B-Instruct-2512IQ2_S24.0B6.97 GiB1.33 GiB8.96 GiB0.04 GiB12±8.3%
Transformed-Journey-24BI1-IQ2_S23.6B6.96 GiB1.33 GiB8.96 GiB0.04 GiB12±8.3%
Magistry-24B-v1.1I1-IQ2_S23.6B6.96 GiB1.33 GiB8.96 GiB0.04 GiB12±8.3%
Mergedonia-AETHER-24B-v1aI1-IQ2_S23.6B6.96 GiB1.33 GiB8.96 GiB0.04 GiB12±8.3%
Mergedonia-AETHER-24B-v1bI1-IQ2_S23.6B6.96 GiB1.33 GiB8.96 GiB0.04 GiB12±8.3%
Slimaki-Tavern-24B-v1.3I1-IQ2_S23.6B6.96 GiB1.33 GiB8.96 GiB0.04 GiB12±8.3%
Maginum-Cydoms-24BI1-IQ2_S23.6B6.96 GiB1.33 GiB8.96 GiB0.04 GiB12±8.3%
Maginum-Cydoms-24B-absolute-heresyI1-IQ2_S23.6B6.96 GiB1.33 GiB8.96 GiB0.04 GiB12±8.3%
Morax-24B-v2IQ2_S23.6B6.96 GiB1.33 GiB8.96 GiB0.04 GiB12±8.3%
Mistral-Small-3.2-24B-Instruct-2506-ultra-uncensored-hereticI1-IQ2_S24.0B6.96 GiB1.33 GiB8.96 GiB0.04 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?
1542 of 2118 indexed open-weight models fit a Apple M4 at 16,384 context with q8_0 KV cache, the largest being Qwen3.6-27B-Heretic2-Thinking at I1-IQ2_XXS. 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.