Apple · apple

Apple M4

Apple M4 has 16 GB of unified memory at 120 GB/s — about 11.16 GiB usable after driver and compositor overhead. 1128 of 2118 indexed models fit at 128K context with q8_0 KV. Note only 12 GB of its 16 GB is allocatable to the GPU.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
16 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 922video 15embedding 25vision language 107audio asr 38audio tts 20image 1

What fits at 128K context

largest quantization that fits, per model · 1128 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Phi-4-mini-instruct-abliteratedQ6_K3.8B2.94 GiB8.50 GiB12.00 GiB0.00 GiB9±8.3%
Phi-4-mini-reasoningQ6_K3.8B2.94 GiB8.50 GiB12.00 GiB0.00 GiB9±8.3%
Phi-4-mini-instructQ6_K3.8B2.94 GiB8.50 GiB12.00 GiB0.00 GiB9±8.3%
Qwen3-Zero-Coder-Reasoning-V2-0.8BI1-IQ2_M816M0.32 GiB11.16 GiB12.00 GiB0.00 GiB9±8.3%
Wan2.1-VACE-14BQ5_K_S17.3B11.41 GiB0.00 GiB12.00 GiB0.00 GiB9±8.3%
zeta-2.1I1-Q2_K_S8.3B2.90 GiB8.50 GiB11.99 GiB0.01 GiB9±8.3%
LFM2-24B-A2BMoEQ3_K_M23.8B10.10 GiB1.33 GiB11.99 GiB0.01 GiB20±37%
Ling-mini-2.0MoEQ4_016.3B8.79 GiB2.66 GiB11.99 GiB0.01 GiB16±37%
gemma-4-12B-it-hereticQ4_112.0B6.89 GiB4.50 GiB11.98 GiB0.02 GiB9±8.3%
Nemotron-3-Embed-8B-BF16Q2_08.0B2.36 GiB9.03 GiB11.98 GiB0.02 GiB9±8.3%
granite-8b-code-instruct-4kI1-IQ1_M8.1B1.83 GiB9.56 GiB11.98 GiB0.02 GiB9±8.3%
granite-8b-code-base-4kI1-IQ1_M8.1B1.83 GiB9.56 GiB11.98 GiB0.02 GiB9±8.3%
salamandra-7b-instruct-2606I1-IQ2_M7.8B2.89 GiB8.50 GiB11.97 GiB0.03 GiB9±8.3%
Ministral-3-8B-Instruct-2512-BF16-abliteratedI1-IQ2_XXS8.9B2.35 GiB9.03 GiB11.97 GiB0.03 GiB9±8.3%
Amaretto-8BI1-IQ2_XXS8.9B2.35 GiB9.03 GiB11.97 GiB0.03 GiB9±8.3%
InternVL3_5-30B-A3BIQ3_XXS30.8B11.38 GiB0.00 GiB11.97 GiB0.03 GiB9±8.3%
Qwen3.6-27B-Heretic2-ThinkingI1-IQ1_M27.4B7.11 GiB4.25 GiB11.97 GiB0.03 GiB9±8.3%
Qwen3.6-27B-Uncensored-AggressiveI1-IQ1_M27.4B7.11 GiB4.25 GiB11.97 GiB0.03 GiB9±8.3%
Qwen-3.5-Opus-GLM-27BI1-IQ1_M26.9B7.11 GiB4.25 GiB11.97 GiB0.03 GiB9±8.3%
Qwen3.6-27B-abliteratedI1-IQ1_M27.4B7.11 GiB4.25 GiB11.97 GiB0.03 GiB9±8.3%
KoQweopus-3.5-27B-experimentalI1-IQ1_M27.8B7.11 GiB4.25 GiB11.97 GiB0.03 GiB9±8.3%
Webcoda-AI-27BI1-IQ1_M27.4B7.11 GiB4.25 GiB11.97 GiB0.03 GiB9±8.3%
Qwen3.5-27B-imabari-v2I1-IQ1_M27.8B7.11 GiB4.25 GiB11.97 GiB0.03 GiB9±8.3%
Qwen3.5-27B-uncensored-heretic-v1I1-IQ1_M27.4B7.11 GiB4.25 GiB11.97 GiB0.03 GiB9±8.3%
Carnice-V2-27bI1-IQ1_M27.4B7.11 GiB4.25 GiB11.97 GiB0.03 GiB9±8.3%
Qwen3.5-Queen-27BI1-IQ1_M27.4B7.11 GiB4.25 GiB11.97 GiB0.03 GiB9±8.3%
GRaPE-2-ProI1-IQ1_M27.8B7.11 GiB4.25 GiB11.97 GiB0.03 GiB9±8.3%
Darwin-28B-REASONI1-IQ1_M26.9B7.11 GiB4.25 GiB11.97 GiB0.03 GiB9±8.3%
Huihui-Qwen3.5-27B-Claude-4.6-Opus-abliteratedI1-IQ1_M27.8B7.11 GiB4.25 GiB11.97 GiB0.03 GiB9±8.3%
Qwen3.5-27B-WebNovel-Writer-zhI1-IQ1_M26.9B7.11 GiB4.25 GiB11.97 GiB0.03 GiB9±8.3%
Qwen3.5-27B_Homebrew-v2I1-IQ1_M27.4B7.11 GiB4.25 GiB11.97 GiB0.03 GiB9±8.3%
CycleGRPO-4BI1-IQ3_XS4.8B1.85 GiB9.56 GiB11.97 GiB0.03 GiB9±8.3%
SuperGemma-4-12b-abliteratedI1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-uncensored-hereticI1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
gemma-4-12B-coder-fable5-composer2.5-v1-uncensored-hereticI1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
gemma-4-12B-it-uncensored-hereticI1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
gemma-4-12B-it-qat-q4_0-unquantized-uncensored-hereticQ4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
Grug-12BI1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
Aura-Medium-v1-BF16I1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
gemma-4-12B-it-Esper4I1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
gemma-4-12B-it-GuardpointI1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
Gemma-4-12B-it-AEON-Abliterated-K4-BF16I1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
gemma-4-12B-it-Tachibana-AgentI1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
gemma-4-12b-marvin-gutenberg-rp-v2I1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
gemma-4-12b-crownelius-writerI1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
Huihui-gemma-4-12B-coder-fable5-composer2.5-v1-abliteratedI1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
gemma-4-12b-asterion-agenticI1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
Huihui-gemma-4-12B-agentic-fable5-abliteratedI1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
g4-12b-it-trismegistusI1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
gemma4-12b-it-asimovI1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
FabGemmaI1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
Huihui-gemma-4-12B-it-qat-q4_0-unquantized-abliteratedI1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
gemma-4-12B-it-abliterated-uncensoredI1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
Gemma-4-12b-it-AbliteratedI1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
gemma-4-12B-Queen-it-qat-q4_0-unquantizedI1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
gemma-4-12B-it-heretic_decensoredI1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
Iris-12B-gemma-4-it-qatI1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
gemma-4-12B-coder-fable5-composer2.5-v1I1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
G4-Starry-Ocean-12BI1-Q4_K_M11.9B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±8.3%
gemma-4-12B-it-QAT-SOMPOA-heresyI1-Q4_K_M12.0B6.87 GiB4.50 GiB11.97 GiB0.03 GiB9±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?
1128 of 2118 indexed open-weight models fit a Apple M4 at 131,072 context with q8_0 KV cache, the largest being Phi-4-mini-instruct-abliterated at Q6_K. That covers text, vision-language, image, video and speech models.
How much usable memory does a Apple M4 actually have?
Its nameplate is 16 GB, but about 11.16 GiB is available to a model once driver and compositor overhead is accounted for, and only 12 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.