Apple · apple

Apple M5 Pro

Apple M5 Pro has 48 GB of unified memory at 307 GB/s — about 33.48 GiB usable after driver and compositor overhead. 2030 of 2118 indexed models fit at 16K context with q4_0 KV. Note only 36 GB of its 48 GB is allocatable to the GPU.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
48 GB
LPDDR5X-9600
Bandwidth
307 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
vision language 183text 1743video 16audio tts 21image 2embedding 26audio asr 39

What fits at 16K context

largest quantization that fits, per model · 2030 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedIQ2_M109B34.56 GiB0.84 GiB35.98 GiB0.02 GiB26±37%
Gemma-2-9B-It-SPPO-Iter3F329.2B34.43 GiB0.95 GiB35.97 GiB0.03 GiB7±8.3%
gemma-2-9b-it-abliteratedF329.2B34.43 GiB0.95 GiB35.97 GiB0.03 GiB7±8.3%
Tiger-Gemma-9B-v1F329.2B34.43 GiB0.95 GiB35.97 GiB0.03 GiB7±8.3%
Step-3.5-Flash-REAP-121B-A11BI1-IQ2_S121B33.40 GiB1.98 GiB35.95 GiB0.05 GiB7±8.3%
Qwen3-Coder-Next-REAMMoEI1-Q4_160.3B35.30 GiB0.11 GiB35.95 GiB0.05 GiB38±37%
Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16MoEIQ4_XS35.1B35.30 GiB0.09 GiB35.95 GiB0.05 GiB36±37%
c4ai-command-r-plus-08-2024Q2_K_S104B34.08 GiB1.13 GiB35.94 GiB0.06 GiB7±8.3%
GLM-4.6VMoEUD-IQ1_S108B34.49 GiB0.81 GiB35.88 GiB0.12 GiB26±37%
Salience-1.5-ProMoEQ8_036.0B35.22 GiB0.09 GiB35.86 GiB0.14 GiB36±37%
Qwable-v1MoEQ8_036.0B35.22 GiB0.09 GiB35.86 GiB0.14 GiB36±37%
T-SearchMoEQ8_036.0B35.22 GiB0.09 GiB35.86 GiB0.14 GiB36±37%
Qwen3.5-35B-A3BMoEQ8_036.0B35.22 GiB0.09 GiB35.86 GiB0.14 GiB36±37%
Behemoth-X-123B-v2IQ2_XS123B33.60 GiB1.55 GiB35.85 GiB0.15 GiB7±8.3%
Mistral-Large-Instruct-2411IQ2_XS123B33.60 GiB1.55 GiB35.85 GiB0.15 GiB7±8.3%
Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-PreservedMoEQ8_035.1B35.21 GiB0.09 GiB35.85 GiB0.15 GiB36±37%
Qwen3.6-35B-A3B-Fable-5-DistillMoEQ8_036.0B35.21 GiB0.09 GiB35.85 GiB0.15 GiB36±37%
Qwable-v2MoEQ8_036.0B35.21 GiB0.09 GiB35.85 GiB0.15 GiB36±37%
Qwen3.6-35B-A3B-YOYO-V2MoEQ8_036.0B35.21 GiB0.09 GiB35.85 GiB0.15 GiB36±37%
Ornith-1.0-35B-FP8-BLOCK-MTPMoEQ8_035.5B35.21 GiB0.09 GiB35.85 GiB0.15 GiB36±37%
fable-coder-35B-A3BMoEQ8_036.0B35.21 GiB0.09 GiB35.85 GiB0.15 GiB36±37%
PINQWEN-3.6-35B-CLEAN-BF16MoEQ8_036.0B35.21 GiB0.09 GiB35.85 GiB0.15 GiB36±37%
UniMath-35B-A3BMoEQ8_036.0B35.21 GiB0.09 GiB35.85 GiB0.15 GiB36±37%
Ornith-1.0-35B-Heretic-MTPMoEQ8_0—35.21 GiB0.09 GiB35.85 GiB0.15 GiB36±37%
Fawen-1.0-35BMoEQ8_036.0B35.21 GiB0.09 GiB35.85 GiB0.15 GiB36±37%
Qwopus3.6-35B-A3B-v1MoEQ8_036.0B35.21 GiB0.09 GiB35.85 GiB0.15 GiB36±37%
CyberStrike-OffSec-35BMoEQ8_035.1B35.21 GiB0.09 GiB35.85 GiB0.15 GiB36±37%
Qwen3.6-35B-A3BMoEQ8_036.0B35.21 GiB0.09 GiB35.85 GiB0.15 GiB36±37%
Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-DistilledMoEQ8_036.0B35.21 GiB0.09 GiB35.85 GiB0.15 GiB36±37%
Qwen3.5-35B-A3B-uncensored-heretic-v2-Native-MTP-PreservedMoEQ8_035.1B35.21 GiB0.09 GiB35.85 GiB0.15 GiB36±37%
Llama-4-Scout-17B-16E-Instruct-abliterated-v2MoEKV unresolvedI1-Q2_K_S109B34.42 GiB0.84 GiB35.84 GiB0.16 GiB26±37%
Hunyuan-A13B-InstructMoEIQ3_M80.4B34.72 GiB0.56 GiB35.82 GiB0.18 GiB7±8.3%
MiniMax-M2.1-REAP-139B-A10BMoEI1-IQ2_XXS139B34.20 GiB1.09 GiB35.82 GiB0.18 GiB26±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-IQ2_XXS139B34.20 GiB1.09 GiB35.82 GiB0.18 GiB26±37%
CodeLlama-70b-Instruct-hfI1-Q3_K_L69.0B33.67 GiB1.41 GiB35.74 GiB0.26 GiB7±8.3%
CodeLlama-70b-Python-hfI1-Q3_K_L69.0B33.67 GiB1.41 GiB35.74 GiB0.26 GiB7±8.3%
Nous-Hermes-Llama2-70bI1-Q3_K_L69.0B33.67 GiB1.41 GiB35.74 GiB0.26 GiB7±8.3%
Midnight-Miqu-70B-v1.5I1-Q3_K_L69.0B33.67 GiB1.41 GiB35.74 GiB0.26 GiB7±8.3%
KafkaLM-70B-German-V0.1Q3_K_L69.0B33.67 GiB1.41 GiB35.74 GiB0.26 GiB7±8.3%
llama2_70b_chat_uncensoredQ3_K_L69.0B33.67 GiB1.41 GiB35.74 GiB0.26 GiB7±8.3%
Xwin-LM-70b-V0.1Q3_K_L69.0B33.67 GiB1.41 GiB35.74 GiB0.26 GiB7±8.3%
Llama-2-70b-chat-hfQ3_K_L69.0B33.67 GiB1.41 GiB35.74 GiB0.26 GiB7±8.3%
dolphin-2.9.1-yi-1.5-34b-hereticQ8_034.4B34.03 GiB1.05 GiB35.72 GiB0.28 GiB7±8.3%
dolphin-2.9.1-yi-1.5-34bQ8_034.4B34.03 GiB1.05 GiB35.72 GiB0.28 GiB7±8.3%
Yi-1.5-34BQ8_034.4B34.03 GiB1.05 GiB35.72 GiB0.28 GiB7±8.3%
Nous-Hermes-2-Yi-34BQ8_034.4B34.03 GiB1.05 GiB35.72 GiB0.28 GiB7±8.3%
Capybara-Tess-Yi-34B-200KQ8_034.4B34.03 GiB1.05 GiB35.72 GiB0.28 GiB7±8.3%
OrionStar-Yi-34B-Chat-LlamaQ8_034.4B34.03 GiB1.05 GiB35.72 GiB0.28 GiB7±8.3%
Nous-Capybara-limarpv3-34BQ8_034.4B34.03 GiB1.05 GiB35.72 GiB0.28 GiB7±8.3%
Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-PreservedQ8_027.4B34.80 GiB0.28 GiB35.69 GiB0.31 GiB7±8.3%
Mistral-Small-4-119B-2603MoEUD-IQ2_M119B34.99 GiB0.10 GiB35.68 GiB0.32 GiB36±37%
Laguna-S-2.1MoEUD-IQ2_M118B34.71 GiB0.25 GiB35.53 GiB0.47 GiB33±37%
deepseek-llm-67b-chatI1-Q3_K_L67.4B33.13 GiB1.67 GiB35.45 GiB0.55 GiB7±8.3%
deepseek-llm-67b-baseI1-Q3_K_L67.4B33.13 GiB1.67 GiB35.45 GiB0.55 GiB7±8.3%
openbuddy-deepseek-67b-v15.3-4kI1-Q3_K_L67.4B33.13 GiB1.67 GiB35.45 GiB0.55 GiB7±8.3%
llama-3.2-3b-instructF163.2B34.37 GiB0.49 GiB35.43 GiB0.57 GiB7±8.3%
llm-surgery-dark-arts-gpt-oss-60b-96a12MoEI1-Q4_160.9B34.76 GiB0.11 GiB35.41 GiB0.59 GiB21±37%
Llama-3_1-Nemotron-51B-InstructQ3_K_M51.5B23.45 GiB11.25 GiB35.39 GiB0.61 GiB7±8.3%
Qwen3-53B-A3B-2507-THINKING-TOTAL-RECALL-v2-MASTER-CODERMoEI1-Q5_K_S53.0B34.03 GiB0.74 GiB35.31 GiB0.69 GiB24±37%
Melody1437-27BQ4_K_M27.8B34.41 GiB0.28 GiB35.31 GiB0.69 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 M5 Pro run?
2030 of 2118 indexed open-weight models fit a Apple M5 Pro at 16,384 context with q4_0 KV cache, the largest being Llama-4-Scout-17B-16E-Instruct at IQ2_M. That covers text, vision-language, image, video and speech models.
How much usable memory does a Apple M5 Pro actually have?
Its nameplate is 48 GB, but about 33.48 GiB is available to a model once driver and compositor overhead is accounted for, and only 36 GB of the pool can be allocated to the GPU at all.
Is a Apple M5 Pro fast for local AI?
Its memory bandwidth is 307 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.