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. 1931 of 2118 indexed models fit at 128K context with q4_0 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
text 1652vision language 175video 16audio asr 39image 2audio tts 21embedding 26

What fits at 128K context

largest quantization that fits, per model · 1931 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
deepseek-coder-33b-instructQ4_K_S33.3B17.64 GiB8.72 GiB26.99 GiB0.01 GiB5±8.3%
WhiteRabbitNeo-33B-v1Q4_K_S33.3B17.64 GiB8.72 GiB26.99 GiB0.01 GiB5±8.3%
Qwen3.5-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-ThinkingI1-Q4_139.5B23.00 GiB3.38 GiB26.99 GiB0.01 GiB5±8.3%
Qwen3.5-40B-RoughHouse-Claude-4.6-Opus-Polar-Deckard-Uncensored-Heretic-ThinkingI1-Q4_139.5B23.00 GiB3.38 GiB26.99 GiB0.01 GiB5±8.3%
deepseek-coder-33b-baseQ4_K_S33.3B17.59 GiB8.72 GiB26.94 GiB0.06 GiB5±8.3%
GLM-Z1-Rumination-32B-0414Q4_K_S33.1B17.72 GiB8.58 GiB26.93 GiB0.07 GiB5±8.3%
Yi-34B-200K-DARE-megamerge-v8IQ4_XS34.4B17.86 GiB8.44 GiB26.93 GiB0.07 GiB5±8.3%
Hy-MT2-30B-A3BMoEQ6_K30.1B23.01 GiB3.38 GiB26.93 GiB0.07 GiB12±37%
Gemma-4-Gembrain-X-Core-31BI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Gemma-4-Gembrain-X-31BI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Gemma-4-31B-Isometry-Fabled-PersonaI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Versipellis-31BI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Gemma4-Gutenberg-31BI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
G4-MeroMero-31B-uncensored-hereticI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Gemma-4-Novelist-31BI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Wanabi-Gemma4-31BI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
G4-Alice-v1.2-31BI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Agares-31B-v1I1-Q5_K_M30.7B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Gemma4-Gutenberg-31B-HereticI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
gemma-4-Ortenzya-The-Creative-Wordsmith-31B-it-uncensored-hereticI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Gemma-4-Gemsicle-31BI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Gemma-4-Gembrain-31B-it-uncensored-hereticI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Melinoe-Gemma4-31B-VL-hereticI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
G4-MeroMero-31BI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Glistening-Gem-31B-v1.0I1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Melinoe-Gemma4-31B-VLI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Gemma-4-31B-Storymaxxed3I1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Gemma-4-AssGuard-31BI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
copywriter-gemma4-31bI1-Q5_K_M32.7B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
gemma-4-31B-heretic-finetuneI1-Q5_K_M30.7B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Gemma-4-Garnet-V2-31B-it-ultra-uncensored-hereticI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
gemma-4-31B-it-Claude-Opus-Distill-v2Q5_K_M32.7B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
gemma-4-31B-it-abliterated-v3I1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Gemma-4-Harmonia-31B-uncensored-hereticQ5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
gemma-4-31B-it-noloopI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Webs-Sejong-31B-v7I1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Lilith-31B-v1.0I1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
JGOS-31B-ThinkI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
gemma-4-31B-MergemaxxedI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
K1-v6-zeroI1-Q5_K_M32.7B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
gemma-4-31B-it-uncensored-hereticQ5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Gemma-4-Queen-31B-it-uncensored-hereticI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Gemma-4-Sphinsikus-Chronist-31BI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
gemma-4-31B-it-hereticI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Gemma4-31B-Finetuned-V2I1-Q5_K_M32.7B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Gemma-4-31B-storymaxxedI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Gemma-4-31B-Fable-5-Agent-DistillQ5_K_M32.7B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
gemma-4-31B-it-uncensoredQ5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Gemma-4-31B-storymaxxed2I1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
gemma-4-31B-it-abliteratedQ5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Gemma-4-Giftige-Blume-31B-v2Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
gemma-4-31B-itQ5_K_M32.7B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
gemma-4-31B-it-Grand-Horror-X-INTENSE-HERETIC-UNCENSORED-ThinkingI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
gemma-4-31B-it-Mystery-Fine-Tune-HERETIC-UNCENSORED-ThinkingI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
gemma-4-31B-it-The-DECKARD-HERETIC-UNCENSORED-ThinkingI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Huihui-gemma-4-31B-it-abliterated-v2I1-Q5_K_M32.7B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Gemma-4-Queen-31B-itI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
gemma-4-31b-it-heretic-araI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Monika-31BI1-Q5_K_M31.3B20.35 GiB5.95 GiB26.93 GiB0.07 GiB5±8.3%
Gemma-4-31B-Fable-CoderI1-Q5_K_M32.7B20.35 GiB5.95 GiB26.93 GiB0.07 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?
1931 of 2118 indexed open-weight models fit a Apple M3 Pro at 131,072 context with q4_0 KV cache, the largest being deepseek-coder-33b-instruct at Q4_K_S. 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.