Apple · apple

Apple M3 Pro

Apple M3 Pro has 36 GB of unified memory at 154 GB/s — about 25.11 GiB usable after driver and compositor overhead. 1841 of 2118 indexed models fit at 64K context with f16 KV. Note only 27 GB of its 36 GB is allocatable to the GPU.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
36 GB
LPDDR5-6400
Bandwidth
154 GB/s
192-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 172text 1566audio asr 39video 16image 1embedding 26audio tts 21

What fits at 64K context

largest quantization that fits, per model · 1841 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
gemma-3-12b-it-abliterated-v2F1611.8B21.92 GiB4.47 GiB26.99 GiB0.01 GiB5±8.3%
gemma-3-12b-it-abliteratedBF1612.2B21.92 GiB4.47 GiB26.99 GiB0.01 GiB5±8.3%
gemma-3-12b-itBF1612.2B21.92 GiB4.47 GiB26.99 GiB0.01 GiB5±8.3%
Qwen3-VL-30B-A3B-ThinkingMoEQ5_K_L31.1B20.43 GiB6.00 GiB26.97 GiB0.03 GiB9±37%
MiroThinker-v1.0-30BMoEQ5_K_L30.5B20.43 GiB6.00 GiB26.97 GiB0.03 GiB9±37%
Qwen3-30B-A3BMoEQ5_K_L30.5B20.43 GiB6.00 GiB26.97 GiB0.03 GiB9±37%
Qwen3-30B-A3B-Instruct-2507MoEQ5_K_L30.5B20.43 GiB6.00 GiB26.97 GiB0.03 GiB9±37%
Qwen3-30B-A3B-Thinking-2507MoEQ5_K_L30.5B20.43 GiB6.00 GiB26.97 GiB0.03 GiB9±37%
Pantheon-Proto-RP-1.8-30B-A3BMoEQ5_K_L30.5B20.43 GiB6.00 GiB26.97 GiB0.03 GiB9±37%
Tongyi-DeepResearch-30B-A3BMoEQ5_K_L30.5B20.43 GiB6.00 GiB26.97 GiB0.03 GiB9±37%
Skyfall-31B-v4.2-hereticI1-Q3_K_S31.4B12.80 GiB13.50 GiB26.97 GiB0.03 GiB5±8.3%
Skyfall-31B-v4.2I1-Q3_K_S31.4B12.80 GiB13.50 GiB26.97 GiB0.03 GiB5±8.3%
MN-GRAND-23.5B-Gutenberg-UNCENSORED-V2-GLM4.7-ThinkingI1-IQ2_XXS23.4B6.12 GiB20.25 GiB26.96 GiB0.04 GiB5±8.3%
granite-4.1-8bBF168.8B16.38 GiB10.00 GiB26.96 GiB0.04 GiB5±8.3%
Noromaid-v0.4-Mixtral-Instruct-8x7b-ZlossMoEIQ3_XS46.7B18.35 GiB8.00 GiB26.94 GiB0.06 GiB5±37%
Nemotron-Cascade-2-30B-A3BMoEQ4_K_L31.6B23.15 GiB3.25 GiB26.93 GiB0.07 GiB12±37%
EXAONE-4.0-32BQ5_K_L32.0B21.44 GiB4.84 GiB26.93 GiB0.07 GiB5±8.3%
dolphin-2.9.2-Phi-3-MediumKV unresolvedQ8_014.0B13.82 GiB12.50 GiB26.92 GiB0.08 GiB5±8.3%
Phi-3-medium-128k-instructQ8_014.0B13.82 GiB12.50 GiB26.92 GiB0.08 GiB5±8.3%
Phi-3-medium-4k-instructQ8_014.0B13.82 GiB12.50 GiB26.92 GiB0.08 GiB5±8.3%
Llama3.2-30B-A3B-II-Dark-Champion-INSTRUCT-Heretic-Abliterated-UncensoredMoEI1-Q3_K_L30.0B14.59 GiB11.75 GiB26.90 GiB0.10 GiB5±37%
medgemma-27b-itI1-Q6_K28.8B20.64 GiB5.61 GiB26.88 GiB0.12 GiB5±8.3%
gemma-3-27b-it-abliterated-refined-visionI1-Q6_K27.4B20.64 GiB5.61 GiB26.88 GiB0.12 GiB5±8.3%
gemma-3-27b-it-abliteratedQ6_K27.4B20.64 GiB5.61 GiB26.88 GiB0.12 GiB5±8.3%
Nidum-Gemma-3-27B-it-UncensoredI1-Q6_K27.4B20.64 GiB5.61 GiB26.88 GiB0.12 GiB5±8.3%
gemma-3-27b-itQ6_K27.4B20.64 GiB5.61 GiB26.88 GiB0.12 GiB5±8.3%
AtomicGPT-gemma3-27bI1-Q6_K27.4B20.64 GiB5.61 GiB26.88 GiB0.12 GiB5±8.3%
Unbound-v1.12.0-27BI1-Q6_K27.4B20.64 GiB5.61 GiB26.88 GiB0.12 GiB5±8.3%
Mira-v1.12-Ties-27BI1-Q6_K27.4B20.64 GiB5.61 GiB26.88 GiB0.12 GiB5±8.3%
Medgamma27BI1-Q6_K27.0B20.64 GiB5.61 GiB26.88 GiB0.12 GiB5±8.3%
medgemma-27b-text-itQ6_K27.0B20.64 GiB5.61 GiB26.88 GiB0.12 GiB5±8.3%
Apriel-1.6-15b-ThinkerQ8_014.9B14.29 GiB12.00 GiB26.88 GiB0.12 GiB5±8.3%
Salience-1.5-FlashMoEQ5_K_M31.1B20.34 GiB6.00 GiB26.88 GiB0.12 GiB9±37%
Qwen3.5-40B-RoughHouse-Claude-4.6-Opus-Polar-Deckard-Uncensored-Heretic-ThinkingI1-IQ4_XS39.5B20.26 GiB6.00 GiB26.88 GiB0.12 GiB5±8.3%
Laguna-S-2.1MoEIQ1_S118B23.15 GiB3.14 GiB26.86 GiB0.14 GiB13±37%
GLM-4.7-FlashMoEQ6_K31.2B23.00 GiB3.30 GiB26.86 GiB0.14 GiB12±37%
Rocinante-XL-16B-v1Q6_K16.1B12.76 GiB13.50 GiB26.85 GiB0.15 GiB5±8.3%
NuExtract-1.5Q4_K_L3.8B2.30 GiB24.00 GiB26.85 GiB0.15 GiB5±8.3%
Phi-3.5-mini-instructQ4_K_L3.8B2.30 GiB24.00 GiB26.85 GiB0.15 GiB5±8.3%
Phi-3-mini-128k-instructQ4_K_L3.8B2.30 GiB24.00 GiB26.85 GiB0.15 GiB5±8.3%
Phi-3.5-mini-instruct_UncensoredQ4_K_L3.8B2.30 GiB24.00 GiB26.85 GiB0.15 GiB5±8.3%
Phi-3-mini-4k-instructQ4_K_L3.8B2.30 GiB24.00 GiB26.85 GiB0.15 GiB5±8.3%
Magistry-24B-v1.1Q5_K_L23.6B16.18 GiB10.00 GiB26.85 GiB0.15 GiB5±8.3%
TildeOpen-30B-Instruct-LVI1-IQ3_XXS30.7B11.21 GiB15.00 GiB26.85 GiB0.15 GiB5±8.3%
Qwen3.5-99BMoEI1-IQ2_XXS99.0B24.77 GiB1.50 GiB26.85 GiB0.15 GiB17±37%
Seed-OSS-36B-Instruct-biprojected-norm-preserving-abliteratedI1-IQ2_XS36.2B10.19 GiB16.00 GiB26.84 GiB0.16 GiB5±8.3%
Seed-OSS-36B-InstructIQ2_XS36.2B10.19 GiB16.00 GiB26.84 GiB0.16 GiB5±8.3%
Hermes-4.3-36B-hereticI1-IQ2_XS36.2B10.19 GiB16.00 GiB26.84 GiB0.16 GiB5±8.3%
Hermes-4.3-36BIQ2_XS36.2B10.19 GiB16.00 GiB26.84 GiB0.16 GiB5±8.3%
EuroLLM-22B-Instruct-2512Q4_K_M22.6B12.72 GiB13.50 GiB26.83 GiB0.17 GiB5±8.3%
WizardCoder-Python-34B-V1.0I1-IQ3_M33.7B14.18 GiB12.00 GiB26.83 GiB0.17 GiB5±8.3%
Phind-CodeLlama-34B-Python-v1I1-IQ3_M33.7B14.18 GiB12.00 GiB26.83 GiB0.17 GiB5±8.3%
Phind-CodeLlama-34B-v2I1-IQ3_M33.7B14.18 GiB12.00 GiB26.83 GiB0.17 GiB5±8.3%
Qwen3-Coder-Next-REAMMoEI1-IQ3_M60.3B24.78 GiB1.50 GiB26.82 GiB0.18 GiB19±37%
grug-27bQ6_K_L27.4B22.20 GiB4.00 GiB26.82 GiB0.18 GiB5±8.3%
Carnice-V2-27bQ6_K_L27.4B22.20 GiB4.00 GiB26.82 GiB0.18 GiB5±8.3%
Fara1.5-27BQ6_K_L27.4B22.20 GiB4.00 GiB26.82 GiB0.18 GiB5±8.3%
octo-netQ4_13.8B2.24 GiB24.00 GiB26.80 GiB0.20 GiB5±8.3%
gemma-2-27b-itIQ4_XS27.2B13.80 GiB12.31 GiB26.79 GiB0.21 GiB5±8.3%
magnum-v4-27bIQ4_XS27.2B13.80 GiB12.31 GiB26.79 GiB0.21 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 M3 Pro run?
1841 of 2118 indexed open-weight models fit a Apple M3 Pro at 65,536 context with f16 KV cache, the largest being gemma-3-12b-it-abliterated-v2 at F16. That covers text, vision-language, image, video and speech models.
How much usable memory does a Apple M3 Pro actually have?
Its nameplate is 36 GB, but about 25.11 GiB is available to a model once driver and compositor overhead is accounted for, and only 27 GB of the pool can be allocated to the GPU at all.
Is a Apple M3 Pro fast for local AI?
Its memory bandwidth is 154 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.