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. 1653 of 2118 indexed models fit at 16K context with q4_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 1431video 12vision language 122audio asr 39embedding 26audio tts 21image 2

What fits at 16K context

largest quantization that fits, per model · 1653 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Lamarck-14B-v0.7I1-IQ4_XS14.8B7.56 GiB0.84 GiB9.00 GiB0.00 GiB11±8.3%
QwenStock-14BI1-IQ4_XS14.8B7.56 GiB0.84 GiB9.00 GiB0.00 GiB11±8.3%
DeepSeek-R1-Distill-Qwen-14B-UncensoredI1-IQ4_XS14.8B7.56 GiB0.84 GiB9.00 GiB0.00 GiB11±8.3%
dolphincoder-starcoder2-15bKV unresolvedI1-IQ4_XS16.0B8.01 GiB0.35 GiB9.00 GiB0.00 GiB12±8.3%
starcoder2-15bKV unresolvedIQ4_XS16.0B8.01 GiB0.35 GiB9.00 GiB0.00 GiB12±8.3%
Wan2.1-T2V-14BQ4_014.3B8.41 GiB0.00 GiB9.00 GiB0.00 GiB11±8.3%
Nexa-AI-4x4B-InstructMoEI1-Q5_K_S12.1B7.80 GiB0.63 GiB8.99 GiB0.01 GiB12±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%
gemma-4-12B-coder-fable5-composer2.5-v1-abliteratedQ4_K_M12.0B7.98 GiB0.41 GiB8.99 GiB0.01 GiB11±8.3%
gemma-4-12B-coder-fable5-composer2.5-v1-sft-v5-abliteratedQ4_K_M12.0B7.98 GiB0.41 GiB8.99 GiB0.01 GiB11±8.3%
DeepSeek-Coder-V2-Lite-InstructMoEIQ4_NL15.7B8.29 GiB0.13 GiB8.99 GiB0.01 GiB34±37%
DeepSeek-Coder-V2-Lite-BaseMoEIQ4_NL15.7B8.29 GiB0.13 GiB8.99 GiB0.01 GiB34±37%
DeepSeek-V2-Lite-ChatMoEIQ4_NL15.7B8.29 GiB0.13 GiB8.99 GiB0.01 GiB34±37%
Ministral-3-14B-Instruct-2512Q4_K_M13.9B7.67 GiB0.70 GiB8.98 GiB0.02 GiB12±8.3%
Ministral-3-14B-Reasoning-2512Q4_K_M13.9B7.67 GiB0.70 GiB8.98 GiB0.02 GiB12±8.3%
Ministral-3-14B-Instruct-2512-BF16-abliteratedI1-Q4_K_M13.9B7.67 GiB0.70 GiB8.98 GiB0.02 GiB12±8.3%
Ministral-3-14B-abliteratedQ4_K_M13.9B7.67 GiB0.70 GiB8.98 GiB0.02 GiB12±8.3%
Ministral-3-14B-Instruct-2512-BF16Q4_K_M13.9B7.67 GiB0.70 GiB8.98 GiB0.02 GiB12±8.3%
Ministral-3-14B-Reasoning-2512-UncensoredI1-Q4_K_M13.9B7.67 GiB0.70 GiB8.98 GiB0.02 GiB12±8.3%
pagestorm-research-preview-14b-full-bookQ4_K_M13.5B7.67 GiB0.70 GiB8.98 GiB0.02 GiB12±8.3%
InternVL3_5-14BQ4_K_M15.1B8.38 GiB0.00 GiB8.98 GiB0.02 GiB11±8.3%
Qwen3-VL-32B-InstructUD-IQ1_S33.4B7.21 GiB1.13 GiB8.97 GiB0.03 GiB12±8.3%
Qwen3-VL-32B-ThinkingUD-IQ1_S33.4B7.21 GiB1.13 GiB8.97 GiB0.03 GiB12±8.3%
Qwen3-32BUD-IQ1_S32.8B7.21 GiB1.13 GiB8.97 GiB0.03 GiB12±8.3%
Qwen3.5-27B-Engineer-Deckard-GeminiI1-IQ2_S27.7B8.08 GiB0.28 GiB8.97 GiB0.03 GiB12±8.3%
Qwen3.5-27B-HERETIC-Polaris-Advanced-Thinking-Alpha-uncensoredI1-IQ2_S27.4B8.08 GiB0.28 GiB8.97 GiB0.03 GiB12±8.3%
Qwen3.5-27B-Deckard-PKD-Heretic-Uncensored-ThinkingI1-IQ2_S27.4B8.08 GiB0.28 GiB8.97 GiB0.03 GiB12±8.3%
Huihui-Qwen3.5-27B-abliteratedI1-IQ2_S27.8B8.08 GiB0.28 GiB8.97 GiB0.03 GiB12±8.3%
Qwen3.5-27B-Unredacted-MAXI1-IQ2_S27.4B8.08 GiB0.28 GiB8.97 GiB0.03 GiB12±8.3%
Qwen3.5-27B-hereticI1-IQ2_S27.4B8.08 GiB0.28 GiB8.97 GiB0.03 GiB12±8.3%
Qwen3.5-27B-DerestrictedI1-IQ2_S27.8B8.08 GiB0.28 GiB8.97 GiB0.03 GiB12±8.3%
Qwen3.5-27B-Claude-4.6-Opus-Reasoning-DistilledI1-IQ2_S27.8B8.08 GiB0.28 GiB8.97 GiB0.03 GiB12±8.3%
HunyuanVideo-1.5Q8_08.3B8.38 GiB0.00 GiB8.97 GiB0.03 GiB12±8.3%
Qwen3.5-21B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-ThinkingI1-IQ3_XXS21.3B8.15 GiB0.21 GiB8.97 GiB0.03 GiB12±8.3%
Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-ThinkingI1-IQ3_XXS21.3B8.15 GiB0.21 GiB8.97 GiB0.03 GiB12±8.3%
SuperGemma-4-12b-abliteratedI1-Q5_K_M12.0B7.96 GiB0.41 GiB8.97 GiB0.03 GiB12±8.3%
gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-uncensored-hereticI1-Q5_K_M12.0B7.96 GiB0.41 GiB8.97 GiB0.03 GiB12±8.3%
gemma-4-12B-coder-fable5-composer2.5-v1-uncensored-hereticI1-Q5_K_M12.0B7.96 GiB0.41 GiB8.97 GiB0.03 GiB12±8.3%
gemma-4-12B-it-uncensored-hereticI1-Q5_K_M12.0B7.96 GiB0.41 GiB8.97 GiB0.03 GiB12±8.3%
Grug-12BI1-Q5_K_M12.0B7.96 GiB0.41 GiB8.97 GiB0.03 GiB12±8.3%
Aura-Medium-v1-BF16I1-Q5_K_M12.0B7.96 GiB0.41 GiB8.97 GiB0.03 GiB12±8.3%
gemma-4-12B-it-Esper4I1-Q5_K_M12.0B7.96 GiB0.41 GiB8.97 GiB0.03 GiB12±8.3%
gemma-4-12B-it-GuardpointI1-Q5_K_M12.0B7.96 GiB0.41 GiB8.97 GiB0.03 GiB12±8.3%
Gemma-4-12B-it-AEON-Abliterated-K4-BF16I1-Q5_K_M12.0B7.96 GiB0.41 GiB8.97 GiB0.03 GiB12±8.3%
gemma-4-12B-it-Tachibana-AgentI1-Q5_K_M12.0B7.96 GiB0.41 GiB8.97 GiB0.03 GiB12±8.3%
gemma-4-12b-marvin-gutenberg-rp-v2I1-Q5_K_M12.0B7.96 GiB0.41 GiB8.97 GiB0.03 GiB12±8.3%
gemma-4-12b-crownelius-writerI1-Q5_K_M12.0B7.96 GiB0.41 GiB8.97 GiB0.03 GiB12±8.3%
Huihui-gemma-4-12B-coder-fable5-composer2.5-v1-abliteratedI1-Q5_K_M12.0B7.96 GiB0.41 GiB8.97 GiB0.03 GiB12±8.3%
gemma-4-12b-asterion-agenticI1-Q5_K_M12.0B7.96 GiB0.41 GiB8.97 GiB0.03 GiB12±8.3%
Huihui-gemma-4-12B-agentic-fable5-abliteratedI1-Q5_K_M12.0B7.96 GiB0.41 GiB8.97 GiB0.03 GiB12±8.3%
g4-12b-it-trismegistusI1-Q5_K_M12.0B7.96 GiB0.41 GiB8.97 GiB0.03 GiB12±8.3%
gemma4-12b-it-asimovI1-Q5_K_M12.0B7.96 GiB0.41 GiB8.97 GiB0.03 GiB12±8.3%
FabGemmaI1-Q5_K_M12.0B7.96 GiB0.41 GiB8.97 GiB0.03 GiB12±8.3%
gemma-4-12B-it-abliterated-uncensoredI1-Q5_K_M12.0B7.96 GiB0.41 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?
1653 of 2118 indexed open-weight models fit a Apple M4 at 16,384 context with q4_0 KV cache, the largest being Lamarck-14B-v0.7 at I1-IQ4_XS. 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.