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. 1973 of 2118 indexed models fit at 64K 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 1693audio tts 21vision language 176video 16image 2audio asr 39embedding 26

What fits at 64K context

largest quantization that fits, per model · 1973 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Salience-1.5-FlashMoEQ6_K_L31.1B24.76 GiB1.69 GiB26.99 GiB0.01 GiB15±37%
Huihui-Qwen3-Coder-Next-abliteratedMoEQ2_K79.7B26.00 GiB0.42 GiB26.97 GiB0.03 GiB26±37%
ALIA-40b-fc-2606I1-Q4_K_M40.4B22.90 GiB3.38 GiB26.94 GiB0.06 GiB5±8.3%
ALIA-40b-instruct-2606I1-Q4_K_M40.4B22.90 GiB3.38 GiB26.94 GiB0.06 GiB5±8.3%
Yi-34B-200K-DARE-megamerge-v8I1-Q5_K_S34.4B22.08 GiB4.22 GiB26.93 GiB0.07 GiB5±8.3%
dolphin-2.9.1-yi-1.5-34b-hereticQ5_K_S34.4B22.08 GiB4.22 GiB26.93 GiB0.07 GiB5±8.3%
dolphin-2.9.1-yi-1.5-34bI1-Q5_K_S34.4B22.08 GiB4.22 GiB26.93 GiB0.07 GiB5±8.3%
OrionStar-Yi-34B-Chat-LlamaI1-Q5_K_S34.4B22.08 GiB4.22 GiB26.93 GiB0.07 GiB5±8.3%
Yi-34B-200K-LlamafiedI1-Q5_K_S34.4B22.08 GiB4.22 GiB26.93 GiB0.07 GiB5±8.3%
Yi-1.5-34BQ5_K_S34.4B22.08 GiB4.22 GiB26.93 GiB0.07 GiB5±8.3%
Nous-Hermes-2-Yi-34BQ5_034.4B22.08 GiB4.22 GiB26.93 GiB0.07 GiB5±8.3%
Merged-RP-Stew-V2-34BI1-Q5_K_S34.4B22.08 GiB4.22 GiB26.93 GiB0.07 GiB5±8.3%
Capybara-Tess-Yi-34B-200KQ5_034.4B22.08 GiB4.22 GiB26.93 GiB0.07 GiB5±8.3%
Nous-Capybara-limarpv3-34BQ5_034.4B22.08 GiB4.22 GiB26.93 GiB0.07 GiB5±8.3%
deepseek-coder-33b-instructQ5_K_M33.3B21.92 GiB4.36 GiB26.91 GiB0.09 GiB5±8.3%
deepseek-coder-33b-baseQ5_K_M33.3B21.92 GiB4.36 GiB26.91 GiB0.09 GiB5±8.3%
WhiteRabbitNeo-33B-v1Q5_K_M33.3B21.92 GiB4.36 GiB26.91 GiB0.09 GiB5±8.3%
Hypernova-60B-2605MoEI1-IQ2_XXS58.7B25.80 GiB0.57 GiB26.91 GiB0.09 GiB20±37%
Qwen3.6-27B-A3B-CoderMoEQ8_026.7B25.99 GiB0.35 GiB26.90 GiB0.10 GiB20±37%
Olmo-3.1-32B-InstructQ6_K_L32.2B24.86 GiB1.36 GiB26.87 GiB0.13 GiB5±8.3%
Olmo-3.1-32B-ThinkQ6_K_L32.2B24.86 GiB1.36 GiB26.87 GiB0.13 GiB5±8.3%
Olmo-3-32B-ThinkQ6_K_L32.2B24.86 GiB1.36 GiB26.87 GiB0.13 GiB5±8.3%
Qwen3-Coder-REAP-25B-A3BMoEQ8_024.9B24.64 GiB1.69 GiB26.87 GiB0.13 GiB14±37%
Noromaid-v0.4-Mixtral-Instruct-8x7b-ZlossMoEIQ4_XS46.7B24.02 GiB2.25 GiB26.86 GiB0.14 GiB8±37%
Ace-Step1.5Q5_K_M160M25.46 GiB0.85 GiB26.85 GiB0.15 GiB5±8.3%
Qwen3-42B-A3B-2507-Thinking-Abliterated-uncensored-TOTAL-RECALL-v2-Medium-MASTER-CODERMoEI1-Q4_K_M42.4B23.94 GiB2.36 GiB26.84 GiB0.16 GiB13±37%
Llama-4-Scout-17B-16E-Instruct-abliterated-v2MoEKV unresolvedI1-IQ1_M109B22.88 GiB3.38 GiB26.84 GiB0.16 GiB12±37%
Llama3.2-30B-A3B-II-Dark-Champion-INSTRUCT-Heretic-Abliterated-UncensoredMoEI1-Q6_K30.0B22.97 GiB3.30 GiB26.83 GiB0.17 GiB10±37%
Noromaid-20b-v0.1.1I1-IQ1_M20.0B4.44 GiB21.80 GiB26.83 GiB0.17 GiB5±8.3%
GLM-Z1-Rumination-32B-0414Q5_K_M33.1B21.90 GiB4.29 GiB26.83 GiB0.17 GiB5±8.3%
c4ai-command-r-plus-08-2024IQ1_S104B21.59 GiB4.50 GiB26.82 GiB0.18 GiB5±8.3%
EuroLLM-22B-Instruct-2512Q8_022.6B22.41 GiB3.80 GiB26.82 GiB0.18 GiB5±8.3%
Devstral-Small-2-24B-Instruct-2512Q8_024.0B23.33 GiB2.81 GiB26.82 GiB0.18 GiB5±8.3%
Magistral-Small-2509Q8_024.0B23.33 GiB2.81 GiB26.82 GiB0.18 GiB5±8.3%
Magistral-Small-2507Q8_023.6B23.33 GiB2.81 GiB26.82 GiB0.18 GiB5±8.3%
Transformed-Journey-24BQ8_023.6B23.33 GiB2.81 GiB26.81 GiB0.19 GiB5±8.3%
Magistry-24B-v1.1Q8_023.6B23.33 GiB2.81 GiB26.81 GiB0.19 GiB5±8.3%
Mergedonia-AETHER-24B-v1aQ8_023.6B23.33 GiB2.81 GiB26.81 GiB0.19 GiB5±8.3%
Mergedonia-AETHER-24B-v1bQ8_023.6B23.33 GiB2.81 GiB26.81 GiB0.19 GiB5±8.3%
Slimaki-Tavern-24B-v1.3Q8_023.6B23.33 GiB2.81 GiB26.81 GiB0.19 GiB5±8.3%
Maginum-Cydoms-24BQ8_023.6B23.33 GiB2.81 GiB26.81 GiB0.19 GiB5±8.3%
Maginum-Cydoms-24B-absolute-heresyQ8_023.6B23.33 GiB2.81 GiB26.81 GiB0.19 GiB5±8.3%
Dolphin3.0-Mistral-24BQ8_023.6B23.33 GiB2.81 GiB26.81 GiB0.19 GiB5±8.3%
Dolphin3.0-R1-Mistral-24BQ8_023.6B23.33 GiB2.81 GiB26.81 GiB0.19 GiB5±8.3%
Cydonia_VistralQ8_023.6B23.33 GiB2.81 GiB26.81 GiB0.19 GiB5±8.3%
Mistral-Small-3.2-24B-Instruct-2506-ultra-uncensored-hereticQ8_024.0B23.33 GiB2.81 GiB26.81 GiB0.19 GiB5±8.3%
Huihui-Mistral-Small-3.2-24B-Instruct-2506-abliterated-llamacppfixedQ8_024.0B23.33 GiB2.81 GiB26.81 GiB0.19 GiB5±8.3%
Dans-PersonalityEngine-V1.2.0-24bQ8_023.6B23.33 GiB2.81 GiB26.81 GiB0.19 GiB5±8.3%
Mistral-Small-3_2-24B-Instruct-2506-antislop.v2Q8_024.0B23.33 GiB2.81 GiB26.81 GiB0.19 GiB5±8.3%
Mistral-Small-3.2-24B-Instruct-2506Q8_024.0B23.33 GiB2.81 GiB26.81 GiB0.19 GiB5±8.3%
Dans-PersonalityEngine-V1.3.0-24bQ8_023.6B23.33 GiB2.81 GiB26.81 GiB0.19 GiB5±8.3%
Devstral-Small-2507Q8_023.6B23.33 GiB2.81 GiB26.81 GiB0.19 GiB5±8.3%
Goetia-24B-v1.1Q8_023.6B23.33 GiB2.81 GiB26.81 GiB0.19 GiB5±8.3%
Devstral-Small-2505Q8_023.6B23.33 GiB2.81 GiB26.81 GiB0.19 GiB5±8.3%
MS3.2-PaintedFantasy-v3-24BQ8_023.6B23.33 GiB2.81 GiB26.81 GiB0.19 GiB5±8.3%
MagiSeek-Pro-V1Q8_023.6B23.33 GiB2.81 GiB26.81 GiB0.19 GiB5±8.3%
Precog-24B-v1Q8_0—23.33 GiB2.81 GiB26.81 GiB0.19 GiB5±8.3%
Cogidonia-v2-24BQ8_023.6B23.33 GiB2.81 GiB26.81 GiB0.19 GiB5±8.3%
Magidonia-24B-v4.3-heretic-v1.2Q8_023.6B23.33 GiB2.81 GiB26.81 GiB0.19 GiB5±8.3%
Magidonia-24B-v4.3Q8_0—23.33 GiB2.81 GiB26.81 GiB0.19 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?
1973 of 2118 indexed open-weight models fit a Apple M3 Pro at 65,536 context with q4_0 KV cache, the largest being Salience-1.5-Flash at Q6_K_L. 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.