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. 1679 of 2118 indexed models fit at 8K 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
vision language 124video 12text 1455audio tts 21embedding 26audio asr 39image 2

What fits at 8K context

largest quantization that fits, per model · 1679 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Muse-Glimmer-30BIQ2_XXS29.8B8.31 GiB0.06 GiB9.00 GiB0.00 GiB12±8.3%
Wan2.1-T2V-14BQ4_014.3B8.41 GiB0.00 GiB9.00 GiB0.00 GiB11±8.3%
Trinity-2-Codestral-22B-v0.2Q2_K_L22.2B7.89 GiB0.49 GiB8.99 GiB0.01 GiB12±8.3%
Mistral-Small-Drummer-22BQ2_K_L22.2B7.89 GiB0.49 GiB8.99 GiB0.01 GiB12±8.3%
Mistral-Small-Instruct-2409Q2_K_L22.2B7.89 GiB0.49 GiB8.99 GiB0.01 GiB12±8.3%
Mistral-Small-22B-ArliAI-RPMax-v1.1Q2_K_L22.2B7.89 GiB0.49 GiB8.99 GiB0.01 GiB12±8.3%
magnum-v4-22bQ2_K_L22.2B7.89 GiB0.49 GiB8.99 GiB0.01 GiB12±8.3%
Qwen3-15B-A2B-BaseMoEQ4_K_S15.6B8.33 GiB0.11 GiB8.99 GiB0.01 GiB41±37%
MythoMax-L2-Kimiko-v2-13bQ3_K_L13.0B6.63 GiB1.76 GiB8.98 GiB0.02 GiB11±8.3%
MythoMax-L2-13bI1-Q3_K_L13.0B6.63 GiB1.76 GiB8.98 GiB0.02 GiB11±8.3%
GPT-NeoX-20B-ErebusI1-IQ2_S20.6B6.02 GiB2.32 GiB8.98 GiB0.02 GiB12±8.3%
Qwen2.5-Coder-7B-InstructQ4_07.6B8.25 GiB0.12 GiB8.98 GiB0.02 GiB12±8.3%
EVA-abliterated-TIES-Qwen2.5-14BI1-IQ4_NL14.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB11±8.3%
Neuron-V1-14B-InstructI1-IQ4_NL14.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB11±8.3%
Ektome-Qwen2.5-Coder-14B-Instruct-PristinelyUncensoredI1-IQ4_NL14.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB11±8.3%
Qwen2.5-14B-Instruct-1M-abliteratedI1-IQ4_NL14.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB11±8.3%
DeepCoder-14B-PreviewIQ4_NL14.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB11±8.3%
Deepseeker-Kunou-Qwen2.5-14bI1-IQ4_NL14.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB11±8.3%
Sugoi-14B-Ultra-HFI1-IQ4_NL14.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB11±8.3%
OpenCodeReasoning-Nemotron-14BIQ4_NL14.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB11±8.3%
DeepSeek-R1-Distill-Qwen-14B-abliterated-v2I1-IQ4_NL14.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB11±8.3%
C1-TachuI1-IQ4_NL14.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB11±8.3%
DeepSeek-R1-Distill-Qwen-14B-abliteratedI1-IQ4_NL14.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB11±8.3%
0x-liteIQ4_NL14.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB11±8.3%
Qwen2.5-Coder-14B-InstructIQ4_NL14.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB11±8.3%
Tessera-4I1-IQ4_NL14.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB11±8.3%
AceReason-Nemotron-14BIQ4_NL14.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB11±8.3%
Tessera-4.1I1-IQ4_NL14.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB11±8.3%
Qwen2.5-14B-Instruct-1MIQ4_NL14.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB11±8.3%
DeepSeek-R1-Distill-Qwen-14BIQ4_NL14.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB11±8.3%
UwU-14B-Math-v0.2I1-IQ4_NL14.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB11±8.3%
oxy-1-smallIQ4_NL14.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB11±8.3%
Impish_QWEN_14B-1MI1-IQ4_NL14.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB11±8.3%
InternVL3_5-14BQ4_K_M15.1B8.38 GiB0.00 GiB8.98 GiB0.02 GiB11±8.3%
Lamarck-14B-v0.7I1-IQ4_NL14.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB11±8.3%
DeepSeek-R1-Distill-Qwen-14B-UncensoredI1-IQ4_NL14.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB11±8.3%
Kepler-8B-Instruct-v2Q4_07.6B8.25 GiB0.12 GiB8.98 GiB0.02 GiB12±8.3%
SuperNova-MediusQ4_014.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB12±8.3%
14B-Qwen2.5-Kunou-v1I1-Q4_014.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB12±8.3%
Qwen2.5-Coder-14B-Instruct-abliteratedQ4_014.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB12±8.3%
Qwen2.5-14B-InstructQ4_014.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB12±8.3%
Qwen2.5-Coder-14BQ4_014.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB12±8.3%
EVA-Qwen2.5-14B-v0.2I1-Q4_014.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB12±8.3%
EVA-Qwen2.5-14B-v0.0I1-Q4_014.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB12±8.3%
EVA-Qwen2.5-14B-v0.1I1-Q4_014.8B7.96 GiB0.42 GiB8.98 GiB0.02 GiB12±8.3%
QwenStock-14BI1-Q4_014.8B7.95 GiB0.42 GiB8.97 GiB0.03 GiB12±8.3%
HunyuanVideo-1.5Q8_08.3B8.38 GiB0.00 GiB8.97 GiB0.03 GiB12±8.3%
Qwen3-14B-Claude-4.5-Opus-High-Reasoning-DistillIQ4_NL14.8B8.01 GiB0.35 GiB8.97 GiB0.03 GiB12±8.3%
Tini-Cybersec-8B-A1BMoEQ8_08.5B8.39 GiB0.03 GiB8.96 GiB0.04 GiB31±37%
LFM2.5-8B-A1B-KO-SFTMoEQ8_08.5B8.39 GiB0.03 GiB8.96 GiB0.04 GiB31±37%
LFM2.5-8B-A1B-SOMPOA-heresyMoEQ8_08.5B8.39 GiB0.03 GiB8.96 GiB0.04 GiB31±37%
Huihui-LFM2.5-8B-A1B-abliteratedMoEQ8_08.5B8.39 GiB0.03 GiB8.96 GiB0.04 GiB31±37%
LFM2.5-8B-A1BMoEQ8_08.5B8.39 GiB0.03 GiB8.96 GiB0.04 GiB31±37%
Supertron2.1-8B-A1BMoEQ8_08.5B8.39 GiB0.03 GiB8.96 GiB0.04 GiB31±37%
LFM2.5-8B-A1B-hereticMoEQ8_08.5B8.39 GiB0.03 GiB8.96 GiB0.04 GiB31±37%
solar-pro-preview-instructKV unresolvedQ2_K22.1B7.65 GiB0.70 GiB8.96 GiB0.04 GiB12±8.3%
gemma-3-12b-it-abliteratedQ5_K_L12.2B8.09 GiB0.27 GiB8.96 GiB0.04 GiB12±8.3%
reka-flash-3.1I1-Q2_K20.9B8.04 GiB0.29 GiB8.95 GiB0.05 GiB12±8.3%
reka-flash-3Q2_K20.9B8.04 GiB0.29 GiB8.95 GiB0.05 GiB12±8.3%
MN-GRAND-23.5B-Gutenberg-UNCENSORED-V2-GLM4.7-ThinkingI1-IQ2_M23.4B7.64 GiB0.71 GiB8.95 GiB0.05 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?
1679 of 2118 indexed open-weight models fit a Apple M4 at 8,192 context with q4_0 KV cache, the largest being Muse-Glimmer-30B at 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.
Apple M4 — what AI models can it run locally? — ossmodeldb