Apple · apple

Apple M2

Apple M2 has 24 GB of unified memory at 102 GB/s — about 16.74 GiB usable after driver and compositor overhead. 1794 of 2118 indexed models fit at 64K context with q8_0 KV. Note only 18 GB of its 24 GB is allocatable to the GPU.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
24 GB
LPDDR5-6400
Bandwidth
102 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 1521vision language 170video 16audio asr 39image 1embedding 26audio tts 21

What fits at 64K context

largest quantization that fits, per model · 1794 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
dolphin-2.6-mixtral-8x7bMoEI1-IQ2_S46.7B13.16 GiB4.25 GiB17.99 GiB0.01 GiB6±37%
xLAM-8x7b-rMoEIQ2_S46.7B13.16 GiB4.25 GiB17.99 GiB0.01 GiB6±37%
Phi-3-medium-4k-instructQ6_K_L14.0B10.74 GiB6.64 GiB17.99 GiB0.01 GiB5±8.3%
Ling-mini-2.0MoEQ8_016.3B16.12 GiB1.33 GiB17.99 GiB0.01 GiB16±37%
Transformed-Journey-24BIQ4_XS23.6B12.00 GiB5.31 GiB17.99 GiB0.01 GiB5±8.3%
Mergedonia-AETHER-24B-v1aIQ4_XS23.6B12.00 GiB5.31 GiB17.99 GiB0.01 GiB5±8.3%
Mergedonia-AETHER-24B-v1bIQ4_XS23.6B12.00 GiB5.31 GiB17.99 GiB0.01 GiB5±8.3%
Slimaki-Tavern-24B-v1.3IQ4_XS23.6B12.00 GiB5.31 GiB17.99 GiB0.01 GiB5±8.3%
Maginum-Cydoms-24BIQ4_XS23.6B12.00 GiB5.31 GiB17.99 GiB0.01 GiB5±8.3%
Maginum-Cydoms-24B-absolute-heresyIQ4_XS23.6B12.00 GiB5.31 GiB17.99 GiB0.01 GiB5±8.3%
Mistral-Small-3.2-24B-Instruct-2506-ultra-uncensored-hereticIQ4_XS24.0B12.00 GiB5.31 GiB17.99 GiB0.01 GiB5±8.3%
Huihui-Mistral-Small-3.2-24B-Instruct-2506-abliterated-llamacppfixedIQ4_XS24.0B12.00 GiB5.31 GiB17.99 GiB0.01 GiB5±8.3%
Mistral-Small-3_2-24B-Instruct-2506-antislop.v2IQ4_XS24.0B12.00 GiB5.31 GiB17.99 GiB0.01 GiB5±8.3%
Goetia-24B-v1.1IQ4_XS23.6B12.00 GiB5.31 GiB17.99 GiB0.01 GiB5±8.3%
MagiSeek-Pro-V1IQ4_XS23.6B12.00 GiB5.31 GiB17.99 GiB0.01 GiB5±8.3%
Cogidonia-v2-24BIQ4_XS23.6B12.00 GiB5.31 GiB17.99 GiB0.01 GiB5±8.3%
Precog-24B-v1IQ4_XS—12.00 GiB5.31 GiB17.99 GiB0.01 GiB5±8.3%
experiment024bIQ4_XS23.6B12.00 GiB5.31 GiB17.99 GiB0.01 GiB5±8.3%
Cydonia-24B-v4.3-absolute-heresyIQ4_XS23.6B12.00 GiB5.31 GiB17.99 GiB0.01 GiB5±8.3%
Cydonia-24B-v4.3-heretic-v2IQ4_XS23.6B12.00 GiB5.31 GiB17.99 GiB0.01 GiB5±8.3%
Cydonia-24B-v4.3-hereticIQ4_XS23.6B12.00 GiB5.31 GiB17.99 GiB0.01 GiB5±8.3%
Cydonia-24B-v4.3-heretic-v4IQ4_XS23.6B12.00 GiB5.31 GiB17.99 GiB0.01 GiB5±8.3%
Journeys-End-24BIQ4_XS23.6B12.00 GiB5.31 GiB17.99 GiB0.01 GiB5±8.3%
Dolphin-Mistral-GLM-4.7-Flash-24B-Venice-Edition-Thinking-UncensoredIQ4_XS23.6B12.00 GiB5.31 GiB17.99 GiB0.01 GiB5±8.3%
WeirdCompound-v1.7-24bIQ4_XS23.6B12.00 GiB5.31 GiB17.99 GiB0.01 GiB5±8.3%
Mistral-Small-24B-Instruct-JbliteratedIQ4_XS23.6B12.00 GiB5.31 GiB17.99 GiB0.01 GiB5±8.3%
grok-oss-Apollyon-24BIQ4_XS23.6B12.00 GiB5.31 GiB17.99 GiB0.01 GiB5±8.3%
grok-oss-Apollyon-24B-hereticIQ4_XS23.6B12.00 GiB5.31 GiB17.99 GiB0.01 GiB5±8.3%
gemma-4-31B-it-qat-q4_0-unquantized-hereticQ2_K_L31.3B11.42 GiB5.94 GiB17.98 GiB0.02 GiB5±8.3%
gemma-2-27b-itIQ3_XS27.2B10.76 GiB6.54 GiB17.98 GiB0.02 GiB5±8.3%
magnum-v4-27bIQ3_XS27.2B10.76 GiB6.54 GiB17.98 GiB0.02 GiB5±8.3%
Wan2.1-VACE-14BQ8_017.3B17.38 GiB0.00 GiB17.97 GiB0.03 GiB5±8.3%
GRM-2.6-Plus-0628Q4_027.8B15.23 GiB2.13 GiB17.96 GiB0.04 GiB5±8.3%
ThinkingCap-Qwen3.6-27BQ4_027.4B15.23 GiB2.13 GiB17.96 GiB0.04 GiB5±8.3%
Tess-4-27BQ4_027.8B15.23 GiB2.13 GiB17.96 GiB0.04 GiB5±8.3%
Qwen3.8-27BIQ4_NL27.8B15.22 GiB2.13 GiB17.95 GiB0.05 GiB5±8.3%
Qwen3.6-27BIQ4_NL27.8B15.22 GiB2.13 GiB17.95 GiB0.05 GiB5±8.3%
InternVL3_5-30B-A3BQ4_K_M30.8B17.35 GiB0.00 GiB17.95 GiB0.05 GiB5±8.3%
Skyfall-31B-v4.2IQ2_S31.4B10.10 GiB7.17 GiB17.94 GiB0.06 GiB5±8.3%
UncensoredLM-DeepSeek-R1-Distill-Qwen-14BQ6_K_L14.2B11.22 GiB6.11 GiB17.93 GiB0.07 GiB5±8.3%
ALIA-40b-fc-2606I1-IQ2_XXS40.4B10.89 GiB6.38 GiB17.93 GiB0.07 GiB5±8.3%
ALIA-40b-instruct-2606I1-IQ2_XXS40.4B10.89 GiB6.38 GiB17.93 GiB0.07 GiB5±8.3%
dolphin-2.9.2-Phi-3-MediumKV unresolvedQ6_K14.0B10.67 GiB6.64 GiB17.92 GiB0.08 GiB5±8.3%
Phi-3-medium-128k-instructQ6_K14.0B10.67 GiB6.64 GiB17.92 GiB0.08 GiB5±8.3%
gemma-7bI1-IQ2_XXS8.5B2.41 GiB14.88 GiB17.91 GiB0.09 GiB5±8.3%
c4ai-command-r-08-2024Q2_K32.3B11.93 GiB5.31 GiB17.91 GiB0.09 GiB5±8.3%
Llama3.2-30B-A3B-II-Dark-Champion-INSTRUCT-Heretic-Abliterated-UncensoredMoEI1-IQ3_XXS30.0B11.09 GiB6.24 GiB17.90 GiB0.10 GiB6±37%
Qwen3-VL-8B-Instruct-HereticI1-Q6_K8.8B12.53 GiB4.78 GiB17.89 GiB0.11 GiB5±8.3%
MiniCPM-V-4_5Q6_K8.7B12.53 GiB4.78 GiB17.89 GiB0.11 GiB5±8.3%
gemma-4-26B-A4B-itMoEQ4_K_M26.5B15.87 GiB1.48 GiB17.89 GiB0.11 GiB5±8.3%
Devstral-Small-2-24B-Instruct-2512IQ4_XS24.0B11.90 GiB5.31 GiB17.88 GiB0.12 GiB5±8.3%
Mistral-Small-3.2-24B-Instruct-2506IQ4_XS24.0B11.90 GiB5.31 GiB17.88 GiB0.12 GiB5±8.3%
Devstral-Small-2507IQ4_XS23.6B11.90 GiB5.31 GiB17.88 GiB0.12 GiB5±8.3%
Devstral-Small-2505IQ4_XS23.6B11.90 GiB5.31 GiB17.88 GiB0.12 GiB5±8.3%
Magistral-Small-2509IQ4_XS24.0B11.90 GiB5.31 GiB17.88 GiB0.12 GiB5±8.3%
Magistral-Small-2507IQ4_XS23.6B11.90 GiB5.31 GiB17.88 GiB0.12 GiB5±8.3%
Mistral-Small-3.1-24B-Instruct-2503IQ4_XS24.0B11.90 GiB5.31 GiB17.88 GiB0.12 GiB5±8.3%
Magistral-Small-2506IQ4_XS23.6B11.90 GiB5.31 GiB17.88 GiB0.12 GiB5±8.3%
Qwen3-Coder-REAP-25B-A3BMoEQ4_K_M24.9B14.15 GiB3.19 GiB17.88 GiB0.12 GiB9±37%
OmniAtlas-Qwen3-30B-A3BI1-Q4_K_M31.7B17.28 GiB0.00 GiB17.88 GiB0.12 GiB5±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 M2 run?
1794 of 2118 indexed open-weight models fit a Apple M2 at 65,536 context with q8_0 KV cache, the largest being dolphin-2.6-mixtral-8x7b at I1-IQ2_S. That covers text, vision-language, image, video and speech models.
How much usable memory does a Apple M2 actually have?
Its nameplate is 24 GB, but about 16.74 GiB is available to a model once driver and compositor overhead is accounted for, and only 18 GB of the pool can be allocated to the GPU at all.
Is a Apple M2 fast for local AI?
Its memory bandwidth is 102 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.