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. 1867 of 2118 indexed models fit at 16K context with f16 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 1592vision language 171video 16embedding 26audio tts 21image 2audio asr 39

What fits at 16K context

largest quantization that fits, per model · 1867 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Nemotron-Cascade-2-30B-A3B-heretic-ara-uncensoredMoEI1-IQ2_XXS31.6B16.65 GiB0.81 GiB18.00 GiB0.00 GiB18±37%
Nemotron-Cascade-2-30B-A3BMoEI1-IQ2_XXS31.6B16.65 GiB0.81 GiB18.00 GiB0.00 GiB18±37%
OLMo-2-1124-13B-InstructQ2_K13.7B4.90 GiB12.50 GiB17.99 GiB0.01 GiB5±8.3%
grug-27bQ4_127.4B16.38 GiB1.00 GiB17.99 GiB0.01 GiB5±8.3%
Carnice-V2-27bQ4_127.4B16.38 GiB1.00 GiB17.99 GiB0.01 GiB5±8.3%
Fara1.5-27BQ4_127.4B16.38 GiB1.00 GiB17.99 GiB0.01 GiB5±8.3%
deepseek-coder-33b-instructIQ3_S33.3B13.49 GiB3.88 GiB17.99 GiB0.01 GiB5±8.3%
GLM-4-32B-0414-Korean-CultureI1-IQ4_XS32.6B16.39 GiB0.95 GiB17.98 GiB0.02 GiB5±8.3%
Muse-Glimmer-30BQ4_K_L29.8B17.05 GiB0.30 GiB17.98 GiB0.02 GiB5±8.3%
Wan2.1-VACE-14BQ8_017.3B17.38 GiB0.00 GiB17.97 GiB0.03 GiB5±8.3%
zeta-2.1F168.3B15.37 GiB2.00 GiB17.96 GiB0.04 GiB5±8.3%
zeta-2BF168.3B15.37 GiB2.00 GiB17.96 GiB0.04 GiB5±8.3%
gemma-4-26B-A4B-it-heretic-ara-v2MoEQ4_K_M25.8B16.51 GiB0.92 GiB17.96 GiB0.04 GiB5±8.3%
Fallen-Gemma3-27B-v1Q4_127.4B15.99 GiB1.38 GiB17.95 GiB0.05 GiB5±8.3%
magnum-v2-32bQ3_K_S32.5B13.30 GiB4.00 GiB17.95 GiB0.05 GiB5±8.3%
Qwen3.8-27BQ4_127.8B16.34 GiB1.00 GiB17.95 GiB0.05 GiB5±8.3%
Qwen3.6-27BQ4_127.8B16.34 GiB1.00 GiB17.95 GiB0.05 GiB5±8.3%
InternVL3_5-30B-A3BQ4_K_M30.8B17.35 GiB0.00 GiB17.95 GiB0.05 GiB5±8.3%
Nous-Hermes-2-Yi-34BQ2_K34.4B13.56 GiB3.75 GiB17.94 GiB0.06 GiB5±8.3%
Capybara-Tess-Yi-34B-200KQ2_K34.4B13.56 GiB3.75 GiB17.94 GiB0.06 GiB5±8.3%
OrionStar-Yi-34B-Chat-LlamaQ2_K34.4B13.56 GiB3.75 GiB17.94 GiB0.06 GiB5±8.3%
Nous-Capybara-limarpv3-34BQ2_K34.4B13.56 GiB3.75 GiB17.94 GiB0.06 GiB5±8.3%
deepseek-coder-33b-baseQ3_K_S33.3B13.43 GiB3.88 GiB17.94 GiB0.06 GiB5±8.3%
WhiteRabbitNeo-33B-v1Q3_K_S33.3B13.43 GiB3.88 GiB17.93 GiB0.07 GiB5±8.3%
llm-jp-4-32b-a3b-thinkingMoEIQ4_XS32.1B16.37 GiB1.00 GiB17.93 GiB0.07 GiB15±37%
Qwen3-16B-A3BMoEQ8_016.0B15.89 GiB1.50 GiB17.93 GiB0.07 GiB11±37%
Seed-OSS-36B-InstructUD-IQ3_XXS36.2B13.27 GiB4.00 GiB17.92 GiB0.08 GiB5±8.3%
gemma-4-26B-A4B-it-qat-q4_0-unquantized-uncensored-hereticMoENVFP425.8B16.46 GiB0.92 GiB17.92 GiB0.08 GiB5±8.3%
Gemma-4-26B-A4B-NVFP4MoENVFP414.4B16.46 GiB0.92 GiB17.92 GiB0.08 GiB5±8.3%
MythoMax-L2-13bI1-IQ3_XXS13.0B4.82 GiB12.50 GiB17.91 GiB0.09 GiB5±8.3%
Qwen3-42B-A3B-2507-Thinking-Abliterated-uncensored-TOTAL-RECALL-v2-Medium-MASTER-CODERMoEI1-IQ3_XXS42.4B15.27 GiB2.09 GiB17.91 GiB0.09 GiB12±37%
medgemma-27b-itI1-Q4_K_M28.8B15.41 GiB1.86 GiB17.90 GiB0.10 GiB5±8.3%
gemma-3-27b-it-abliterated-refined-visionI1-Q4_K_M27.4B15.41 GiB1.86 GiB17.90 GiB0.10 GiB5±8.3%
gemma-3-27b-it-abliteratedQ4_K_M27.4B15.41 GiB1.86 GiB17.90 GiB0.10 GiB5±8.3%
Nidum-Gemma-3-27B-it-UncensoredI1-Q4_K_M27.4B15.41 GiB1.86 GiB17.90 GiB0.10 GiB5±8.3%
gemma-3-27b-itQ4_K_M27.4B15.41 GiB1.86 GiB17.90 GiB0.10 GiB5±8.3%
AtomicGPT-gemma3-27bI1-Q4_K_M27.4B15.41 GiB1.86 GiB17.90 GiB0.10 GiB5±8.3%
Unbound-v1.12.0-27BI1-Q4_K_M27.4B15.41 GiB1.86 GiB17.90 GiB0.10 GiB5±8.3%
Mira-v1.12-Ties-27BI1-Q4_K_M27.4B15.41 GiB1.86 GiB17.90 GiB0.10 GiB5±8.3%
Medgamma27BI1-Q4_K_M27.0B15.41 GiB1.86 GiB17.90 GiB0.10 GiB5±8.3%
medgemma-27b-text-itQ4_K_M27.0B15.41 GiB1.86 GiB17.90 GiB0.10 GiB5±8.3%
gemma-2-27b-itQ3_K_L27.2B13.52 GiB3.68 GiB17.89 GiB0.11 GiB5±8.3%
magnum-v4-27bQ3_K_L27.2B13.52 GiB3.68 GiB17.89 GiB0.11 GiB5±8.3%
Ornith-1.0-35B-uncensored-hereticMoEQ3_K_L35.1B17.02 GiB0.31 GiB17.89 GiB0.11 GiB24±37%
Apriel-1.6-15b-ThinkerQ8_014.9B14.29 GiB3.00 GiB17.88 GiB0.12 GiB5±8.3%
OmniAtlas-Qwen3-30B-A3BI1-Q4_K_M31.7B17.28 GiB0.00 GiB17.88 GiB0.12 GiB5±8.3%
Qwen3-Omni-30B-A3B-InstructQ4_K_M35.3B17.28 GiB0.00 GiB17.88 GiB0.12 GiB5±8.3%
Qwen3-Omni-30B-A3B-CaptionerI1-Q4_K_M31.7B17.28 GiB0.00 GiB17.88 GiB0.12 GiB5±8.3%
Qwen3-Omni-30B-A3B-ThinkingQ4_K_M31.7B17.28 GiB0.00 GiB17.88 GiB0.12 GiB5±8.3%
Qwen3-48B-A4B-Savant-Commander-Distill-12X-Closed-Open-Heretic-UncensoredMoEI1-Q3_K_M33.6B15.06 GiB2.25 GiB17.87 GiB0.13 GiB9±37%
Qwen3-Coder-Next-Opus-4.6-Reasoning-DistilledMoEIQ1_M—16.95 GiB0.38 GiB17.87 GiB0.13 GiB25±37%
Olmo-3-7B-InstructBF167.3B13.60 GiB3.69 GiB17.86 GiB0.14 GiB5±8.3%
Olmo-3-7B-ThinkBF167.3B13.60 GiB3.69 GiB17.86 GiB0.14 GiB5±8.3%
granite-8b-code-instruct-4kF168.1B15.01 GiB2.25 GiB17.85 GiB0.15 GiB5±8.3%
granite-8b-code-base-4kF168.1B15.01 GiB2.25 GiB17.85 GiB0.15 GiB5±8.3%
Qwen3.6-35B-A3B-uncensored-hereticMoEQ3_K_L35.1B16.97 GiB0.31 GiB17.83 GiB0.17 GiB24±37%
Nex-N2-mini-ultra-uncensored-hereticMoEQ3_K_L35.1B16.97 GiB0.31 GiB17.83 GiB0.17 GiB24±37%
WizardCoder-Python-34B-V1.0I1-IQ3_M33.7B14.18 GiB3.00 GiB17.83 GiB0.17 GiB5±8.3%
Phind-CodeLlama-34B-Python-v1I1-IQ3_M33.7B14.18 GiB3.00 GiB17.83 GiB0.17 GiB5±8.3%
Phind-CodeLlama-34B-v2I1-IQ3_M33.7B14.18 GiB3.00 GiB17.83 GiB0.17 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?
1867 of 2118 indexed open-weight models fit a Apple M2 at 16,384 context with f16 KV cache, the largest being Nemotron-Cascade-2-30B-A3B-heretic-ara-uncensored at I1-IQ2_XXS. 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.