Apple · apple

Apple M2 Pro

Apple M2 Pro has 32 GB of unified memory at 205 GB/s — about 22.32 GiB usable after driver and compositor overhead. 1883 of 2118 indexed models fit at 32K context with f16 KV. Note only 24 GB of its 32 GB is allocatable to the GPU.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
32 GB
LPDDR5-6400
Bandwidth
205 GB/s
256-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
audio asr 39text 1606vision language 173image 2audio tts 21video 16embedding 26

What fits at 32K context

largest quantization that fits, per model · 1883 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Voxtral-Small-24B-2507Q6_K_L24.3B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Devstral-Small-2-24B-Instruct-2512Q6_K_L24.0B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Dolphin3.0-R1-Mistral-24BQ6_K_L23.6B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Dolphin3.0-Mistral-24BQ6_K_L23.6B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Cydonia_VistralQ6_K_L23.6B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Dans-PersonalityEngine-V1.2.0-24bQ6_K_L23.6B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Dans-PersonalityEngine-V1.3.0-24bQ6_K_L23.6B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Devstral-Small-2505Q6_K_L23.6B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Mistral-Small-3.2-24B-Instruct-2506Q6_K_L24.0B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
MS3.2-PaintedFantasy-v3-24BQ6_K_L23.6B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Precog-24B-v1Q6_K_L—18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Magidonia-24B-v4.3Q6_K_L—18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Magidonia-24B-v4.2.0Q6_K_L23.6B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
MS-2501-DPE-QwQify-v0.1-24BQ6_K_L23.6B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
sarvam-mQ6_K_L23.6B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Magistral-Small-2506Q6_K_L23.6B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Cydonia-24B-v4.1Q6_K_L23.6B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Cydonia-24B-v4Q6_K_L23.6B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Mistral-Small-3.1-24B-Instruct-2503Q6_K_L24.0B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Cydonia-24B-v4.3Q6_K_L23.6B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Cydonia-24B-v4.2.0Q6_K_L23.6B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Mistral-Small-24B-Instruct-2501-abliteratedQ6_K_L23.6B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Dolphin-Mistral-24B-Venice-EditionQ6_K_L24.0B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Magistral-Small-2509-VisionQ6_K_M24.0B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Mistral-Small-24B-Instruct-2501Q6_K_L23.6B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Hearthfire-24BQ6_K_L23.6B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Mistral-Small-24B-ArliAI-RPMax-v1.4Q6_K_L23.6B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Codex-24B-Small-3.2Q6_K_L23.6B18.32 GiB5.00 GiB23.99 GiB0.01 GiB7±8.3%
Gemma-4-Dark-Gemistry-31BQ4_032.7B17.18 GiB6.17 GiB23.98 GiB0.02 GiB7±8.3%
InternVL3_5-30B-A3BQ6_K30.8B23.38 GiB0.00 GiB23.98 GiB0.02 GiB7±8.3%
Qwen3-VL-30B-A3B-ThinkingMoEQ5_K_L31.1B20.43 GiB3.00 GiB23.97 GiB0.03 GiB17±37%
MiroThinker-v1.0-30BMoEQ5_K_L30.5B20.43 GiB3.00 GiB23.97 GiB0.03 GiB17±37%
Qwen3-30B-A3BMoEQ5_K_L30.5B20.43 GiB3.00 GiB23.97 GiB0.03 GiB17±37%
Qwen3-30B-A3B-Instruct-2507MoEQ5_K_L30.5B20.43 GiB3.00 GiB23.97 GiB0.03 GiB17±37%
Qwen3-30B-A3B-Thinking-2507MoEQ5_K_L30.5B20.43 GiB3.00 GiB23.97 GiB0.03 GiB17±37%
Pantheon-Proto-RP-1.8-30B-A3BMoEQ5_K_L30.5B20.43 GiB3.00 GiB23.97 GiB0.03 GiB17±37%
Tongyi-DeepResearch-30B-A3BMoEQ5_K_L30.5B20.43 GiB3.00 GiB23.97 GiB0.03 GiB17±37%
Qwen3.6-27B-Fable-5-ExperimentalQ6_K27.8B21.36 GiB2.00 GiB23.97 GiB0.03 GiB7±8.3%
OmniAtlas-Qwen3-30B-A3BI1-Q6_K31.7B23.37 GiB0.00 GiB23.96 GiB0.04 GiB7±8.3%
Qwen3-Omni-30B-A3B-CaptionerI1-Q6_K31.7B23.37 GiB0.00 GiB23.96 GiB0.04 GiB7±8.3%
gemma-4-31B-itQ4_031.3B17.16 GiB6.17 GiB23.96 GiB0.04 GiB7±8.3%
codegeex4-all-9bIQ2_XS9.4B3.36 GiB20.00 GiB23.96 GiB0.04 GiB7±8.3%
glm-4-9b-chatIQ2_XS9.4B3.36 GiB20.00 GiB23.96 GiB0.04 GiB7±8.3%
spoomplesmaxx-v2.1-30BI1-Q4_028.9B15.29 GiB8.00 GiB23.95 GiB0.05 GiB7±8.3%
Huihui-granite-4.1-30b-abliteratedI1-Q4_028.9B15.29 GiB8.00 GiB23.95 GiB0.05 GiB7±8.3%
granite-4.1-30b-hereticI1-Q4_028.9B15.29 GiB8.00 GiB23.95 GiB0.05 GiB7±8.3%
Nemotron-Labs-Audex-30B-A3BQ4_K_L32.0B23.35 GiB0.00 GiB23.95 GiB0.05 GiB7±8.3%
MN-GRAND-23.5B-Gutenberg-UNCENSORED-V2-GLM4.7-ThinkingI1-Q4_K_M23.4B13.22 GiB10.13 GiB23.94 GiB0.06 GiB7±8.3%
Pantheon-Reasoning-26B-A4B-1.1MoEQ6_K26.5B21.86 GiB1.54 GiB23.94 GiB0.06 GiB7±8.3%
ERNIE-21B-A3B-Claude-4.5-High-OPUS-ThinkingQ8_021.8B21.61 GiB1.75 GiB23.93 GiB0.07 GiB7±8.3%
ERNIE-4.5-21B-A3B-ThinkingQ8_021.8B21.61 GiB1.75 GiB23.93 GiB0.07 GiB7±8.3%
ERNIE-4.5-21B-A3B-PTQ8_021.9B21.61 GiB1.75 GiB23.93 GiB0.07 GiB7±8.3%
Qwen3.8-27BQ6_K27.8B21.31 GiB2.00 GiB23.93 GiB0.07 GiB7±8.3%
Qwen3.6-27BQ6_K27.8B21.31 GiB2.00 GiB23.93 GiB0.07 GiB7±8.3%
Qwen3-53B-A3B-2507-THINKING-TOTAL-RECALL-v2-MASTER-CODERMoEI1-Q2_K53.0B18.10 GiB5.25 GiB23.89 GiB0.11 GiB12±37%
granite-4.1-30bQ4_028.9B15.23 GiB8.00 GiB23.88 GiB0.12 GiB7±8.3%
Salience-1.5-FlashMoEQ5_K_M31.1B20.34 GiB3.00 GiB23.88 GiB0.12 GiB17±37%
Huihui-GLM-4.7-Flash-abliterated-57BMoEI1-Q2_K57.3B19.11 GiB4.18 GiB23.88 GiB0.12 GiB14±37%
Qwen3.5-40B-RoughHouse-Claude-4.6-Opus-Polar-Deckard-Uncensored-Heretic-ThinkingI1-IQ4_XS39.5B20.26 GiB3.00 GiB23.88 GiB0.12 GiB7±8.3%
Fallen-Gemma3-27B-v1Q6_K_L27.4B20.96 GiB2.31 GiB23.87 GiB0.13 GiB7±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 Pro run?
1883 of 2118 indexed open-weight models fit a Apple M2 Pro at 32,768 context with f16 KV cache, the largest being Voxtral-Small-24B-2507 at Q6_K_L. That covers text, vision-language, image, video and speech models.
How much usable memory does a Apple M2 Pro actually have?
Its nameplate is 32 GB, but about 22.32 GiB is available to a model once driver and compositor overhead is accounted for, and only 24 GB of the pool can be allocated to the GPU at all.
Is a Apple M2 Pro fast for local AI?
Its memory bandwidth is 205 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.