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. 1953 of 2118 indexed models fit at 4K context with q4_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 1676vision language 173video 16image 2audio asr 39audio tts 21embedding 26

What fits at 4K context

largest quantization that fits, per model · 1953 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Tongyi-DeepResearch-30B-A3BMoEQ4_K_M30.5B17.35 GiB0.11 GiB18.00 GiB0.00 GiB20±37%
Tess-4-27BQ4_K_M27.8B17.30 GiB0.07 GiB17.99 GiB0.01 GiB5±8.3%
Olmo-3.1-32B-InstructIQ4_NL32.2B17.06 GiB0.28 GiB17.98 GiB0.02 GiB5±8.3%
Olmo-3.1-32B-ThinkIQ4_NL32.2B17.06 GiB0.28 GiB17.98 GiB0.02 GiB5±8.3%
Olmo-3-32B-ThinkIQ4_NL32.2B17.06 GiB0.28 GiB17.98 GiB0.02 GiB5±8.3%
Gemma4-Gutenberg-31BQ4_031.3B16.83 GiB0.51 GiB17.97 GiB0.03 GiB5±8.3%
Gemma4-Gutenberg-31B-HereticQ4_031.3B16.83 GiB0.51 GiB17.97 GiB0.03 GiB5±8.3%
Equinox-31BQ4_031.3B16.83 GiB0.51 GiB17.97 GiB0.03 GiB5±8.3%
gemma-4-31B-it-SDFT-Heretic-RPQ4_030.7B16.83 GiB0.51 GiB17.97 GiB0.03 GiB5±8.3%
Wan2.1-VACE-14BQ8_017.3B17.38 GiB0.00 GiB17.97 GiB0.03 GiB5±8.3%
Llama-3_1-Nemotron-51B-InstructIQ2_XS51.5B14.46 GiB2.81 GiB17.96 GiB0.04 GiB5±8.3%
EXAONE-4.0-32BQ4_K_S32.0B17.03 GiB0.28 GiB17.96 GiB0.04 GiB5±8.3%
Trinity-MiniMoEQ5_K_M26.1B17.36 GiB0.05 GiB17.95 GiB0.05 GiB21±37%
Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-PreservedMoEQ3_K_L35.1B17.37 GiB0.02 GiB17.95 GiB0.05 GiB27±37%
InternVL3_5-30B-A3BQ4_K_M30.8B17.35 GiB0.00 GiB17.95 GiB0.05 GiB5±8.3%
Qwen3.5-35B-A3B-uncensored-heretic-v2-Native-MTP-PreservedMoEQ3_K_L35.1B17.37 GiB0.02 GiB17.95 GiB0.05 GiB27±37%
Trinity-2-Codestral-22B-v0.2Q6_K_L22.2B17.09 GiB0.25 GiB17.95 GiB0.05 GiB5±8.3%
Mistral-Small-Drummer-22BQ6_K_L22.2B17.09 GiB0.25 GiB17.95 GiB0.05 GiB5±8.3%
Mistral-Small-Instruct-2409Q6_K_L22.2B17.09 GiB0.25 GiB17.95 GiB0.05 GiB5±8.3%
Mistral-Small-22B-ArliAI-RPMax-v1.1Q6_K_L22.2B17.09 GiB0.25 GiB17.95 GiB0.05 GiB5±8.3%
magnum-v4-22bQ6_K_L22.2B17.09 GiB0.25 GiB17.95 GiB0.05 GiB5±8.3%
Huihui-Qwen3.5-35B-A3B-abliteratedMoEI1-IQ4_XS36.0B17.37 GiB0.02 GiB17.95 GiB0.05 GiB27±37%
Qwen3.5-35B-A3B-BaseMoEI1-IQ4_XS36.0B17.37 GiB0.02 GiB17.95 GiB0.05 GiB27±37%
Qwen3.5-35B-A3B-Claude-4.6-Opus-Reasoning-DistilledMoEI1-IQ4_XS36.0B17.37 GiB0.02 GiB17.95 GiB0.05 GiB27±37%
reka-flash-3.1I1-Q6_K20.9B17.17 GiB0.15 GiB17.94 GiB0.06 GiB5±8.3%
reka-flash-3Q6_K20.9B17.17 GiB0.15 GiB17.94 GiB0.06 GiB5±8.3%
Salience-1.5-FlashMoEI1-Q4_K_M31.1B17.28 GiB0.11 GiB17.93 GiB0.07 GiB20±37%
Huihui-Qwen3-VL-30B-A3B-Instruct-abliteratedMoEI1-Q4_K_M31.1B17.28 GiB0.11 GiB17.93 GiB0.07 GiB20±37%
Qwen3-VL-30B-A3B-InstructMoEQ4_K_M31.1B17.28 GiB0.11 GiB17.93 GiB0.07 GiB20±37%
Qwen3-30B-A3B-Gemini-Pro-High-Reasoning-2507-ABLITERATED-UNCENSOREDMoEI1-Q4_K_M30.5B17.28 GiB0.11 GiB17.93 GiB0.07 GiB20±37%
MiroThinker-v1.0-30BMoEI1-Q4_K_M30.5B17.28 GiB0.11 GiB17.93 GiB0.07 GiB20±37%
Qwen3-30B-A3B-YOYO-V5MoEI1-Q4_K_M30.5B17.28 GiB0.11 GiB17.93 GiB0.07 GiB20±37%
Qwen3-30B-A3B-Thinking-2507-Claude-4.5-Sonnet-High-Reasoning-DistillMoEI1-Q4_K_M30.5B17.28 GiB0.11 GiB17.93 GiB0.07 GiB20±37%
Huihui-Qwen3-30B-A3B-Thinking-2507-abliteratedMoEI1-Q4_K_M30.5B17.28 GiB0.11 GiB17.93 GiB0.07 GiB20±37%
Huihui-Qwen3-30B-A3B-Instruct-2507-abliteratedMoEI1-Q4_K_M30.5B17.28 GiB0.11 GiB17.93 GiB0.07 GiB20±37%
Qwen3-30B-A3B-abliterated-eroticMoEI1-Q4_K_M30.5B17.28 GiB0.11 GiB17.93 GiB0.07 GiB20±37%
Qwen3-30B-A3B-abliteratedMoEQ4_K_M30.5B17.28 GiB0.11 GiB17.93 GiB0.07 GiB20±37%
Huihui-Qwen3-Coder-30B-A3B-Instruct-abliteratedMoEI1-Q4_K_M30.5B17.28 GiB0.11 GiB17.93 GiB0.07 GiB20±37%
Qwen3-Coder-30B-A3B-Instruct-RTPurboMoEI1-Q4_K_M30.5B17.28 GiB0.11 GiB17.93 GiB0.07 GiB20±37%
gemma-4-31B-itIQ4_NL31.3B16.79 GiB0.51 GiB17.92 GiB0.08 GiB5±8.3%
HarmonicHarlequin_v5-20BI1-Q3_K_M33.3B15.04 GiB2.29 GiB17.91 GiB0.09 GiB5±8.3%
grug-27bQ4_K_L27.4B17.21 GiB0.07 GiB17.89 GiB0.11 GiB5±8.3%
Carnice-V2-27bQ4_K_L27.4B17.21 GiB0.07 GiB17.89 GiB0.11 GiB5±8.3%
Fara1.5-27BQ4_K_L27.4B17.21 GiB0.07 GiB17.89 GiB0.11 GiB5±8.3%
Goetia-26B-A4B-v1.4MoEI1-Q5_K_S26.0B17.22 GiB0.13 GiB17.89 GiB0.11 GiB5±8.3%
G4-Moonlight-Dusk-26B-A4B-hereticMoEI1-Q5_K_S26.5B17.22 GiB0.13 GiB17.89 GiB0.11 GiB5±8.3%
Pantheon-Reasoning-26B-A4B-1.1-hereticMoEI1-Q5_K_S26.5B17.22 GiB0.13 GiB17.89 GiB0.11 GiB5±8.3%
G4-Moonlight-Dusk-26B-A4BMoEI1-Q5_K_S26.5B17.22 GiB0.13 GiB17.89 GiB0.11 GiB5±8.3%
Chimera-X-26B-A4BMoEI1-Q5_K_S26.5B17.22 GiB0.13 GiB17.89 GiB0.11 GiB5±8.3%
Pantheon-Reasoning-26B-A4B-1.1MoEI1-Q5_K_S26.5B17.22 GiB0.13 GiB17.89 GiB0.11 GiB5±8.3%
Gemma-4-26B-A4B-StyleTune-V2MoEI1-Q5_K_S26.5B17.22 GiB0.13 GiB17.89 GiB0.11 GiB5±8.3%
Gemma-4-26B-A4B-StyleTuneMoEI1-Q5_K_S26.5B17.22 GiB0.13 GiB17.89 GiB0.11 GiB5±8.3%
gemma-4-26b-a4b-heretic-styletune-v2-headMoEI1-Q5_K_S25.8B17.22 GiB0.13 GiB17.89 GiB0.11 GiB5±8.3%
gemma-4-E2B-it-Uncensored-MAXF325.1B17.33 GiB0.02 GiB17.89 GiB0.11 GiB5±8.3%
GLM-4.7-Flash-hereticMoEQ4_K_M29.9B17.27 GiB0.06 GiB17.88 GiB0.12 GiB21±37%
Huihui-GLM-4.7-Flash-abliterated-57BMoEI1-IQ2_M57.3B17.14 GiB0.15 GiB17.88 GiB0.12 GiB20±37%
Delphi-25B-SimpleRL-MathI1-Q5_K_S25.0B16.08 GiB1.18 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%
●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?
1953 of 2118 indexed open-weight models fit a Apple M2 at 4,096 context with q4_0 KV cache, the largest being Tongyi-DeepResearch-30B-A3B at Q4_K_M. 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.