Apple · apple

Apple M2 Ultra

Apple M2 Ultra has 192 GB of unified memory at 819 GB/s — about 133.92 GiB usable after driver and compositor overhead. 2092 of 2118 indexed models fit at 128K context with q8_0 KV. Note only 144 GB of its 192 GB is allocatable to the GPU.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
192 GB
LPDDR5-6400
Bandwidth
819 GB/s
1024-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 1797vision language 191image 2audio tts 21audio asr 39video 16embedding 26

What fits at 128K context

largest quantization that fits, per model · 2092 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
dots.llm1.instMoEQ4_0143B77.44 GiB65.88 GiB143.89 GiB0.11 GiB5±37%
DeepSeek-V3-0324MoEIQ1_M685B138.66 GiB4.56 GiB143.84 GiB0.16 GiB21±37%
r1-1776MoEIQ1_M671B138.66 GiB4.56 GiB143.84 GiB0.16 GiB21±37%
DeepSeek-R1MoEIQ1_M685B138.66 GiB4.56 GiB143.84 GiB0.16 GiB21±37%
GLM-4.6-Derestricted-v3MoEQ2_K357B118.51 GiB24.44 GiB143.54 GiB0.46 GiB10±37%
Step-3.7-FlashQ4_1201B116.67 GiB26.05 GiB143.30 GiB0.70 GiB5±8.3%
GLM-4.7-REAP-218B-A32BMoEQ4_0218B117.58 GiB24.44 GiB142.60 GiB1.40 GiB9±37%
Qwen3.5-REAP-212B-A17BMoEQ5_K_M212B139.97 GiB1.99 GiB142.56 GiB1.44 GiB23±37%
DeepSeek-R1-0528MoEIQ1_M685B137.32 GiB4.56 GiB142.51 GiB1.49 GiB21±37%
DeepSeek-V3.1MoEIQ1_M685B137.32 GiB4.56 GiB142.51 GiB1.49 GiB21±37%
Qwen3.5-397B-A17BMoEUD-IQ3_XXS403B139.55 GiB1.99 GiB142.14 GiB1.86 GiB26±37%
step-3.5-flashQ4_1199B115.15 GiB26.05 GiB141.78 GiB2.22 GiB5±8.3%
MiniMax-M2.7MoEQ4_K_S229B124.65 GiB16.47 GiB141.66 GiB2.34 GiB13±37%
MiniMax-M2.1MoEQ4_K_S229B124.56 GiB16.47 GiB141.57 GiB2.43 GiB13±37%
MiniMax-M2MoEQ4_K_S229B124.56 GiB16.47 GiB141.57 GiB2.43 GiB13±37%
Qwen3-Coder-480B-A35B-InstructMoEIQ2_S480B124.42 GiB16.47 GiB141.47 GiB2.53 GiB13±37%
Qwen3-235B-A22B-Instruct-2507MoEQ4_K_S235B128.25 GiB12.48 GiB141.31 GiB2.69 GiB13±37%
Qwen3-235B-A22B-Thinking-2507MoEQ4_K_S235B128.25 GiB12.48 GiB141.31 GiB2.69 GiB13±37%
Qwen3.5-REAP-262B-A17BMoEQ4_K_S262B138.50 GiB1.99 GiB141.09 GiB2.91 GiB24±37%
MiMo-V2.5MoEKV unresolvedQ3_K_M311B132.47 GiB7.97 GiB141.04 GiB2.96 GiB18±37%
Qwen3-Coder-REAP-363B-A35BMoEQ2_K_L363B123.36 GiB16.47 GiB140.41 GiB3.59 GiB12±37%
Solar-Open2-250BMoEIQ4_XS250B126.88 GiB12.75 GiB140.21 GiB3.79 GiB16±37%
Trinity-Large-PreviewMoEQ2_K_L399B135.17 GiB4.40 GiB140.15 GiB3.85 GiB24±37%
Ornith-1.0-397BMoEUD-IQ3_XXS397B137.46 GiB1.99 GiB140.05 GiB3.95 GiB26±37%
Trinity-Large-TrueBaseMoEI1-Q2_K399B134.81 GiB4.40 GiB139.79 GiB4.21 GiB24±37%
QwQ-32BBF1632.8B122.07 GiB17.00 GiB139.72 GiB4.28 GiB5±8.3%
Hy3MoEIQ3_XXS299B117.43 GiB21.25 GiB139.27 GiB4.73 GiB11±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedUD-IQ2_XXS402B125.91 GiB12.75 GiB139.24 GiB4.76 GiB17±37%
MiMo-V2-FlashMoEKV unresolvedIQ3_M310B130.65 GiB7.97 GiB139.22 GiB4.78 GiB18±37%
GLM-4.5MoEUD-IQ2_M358B114.03 GiB24.44 GiB139.06 GiB4.94 GiB10±37%
GLM-4.7MoEUD-IQ2_M358B114.03 GiB24.44 GiB139.06 GiB4.94 GiB10±37%
grok-2MoEQ3_K_L270B121.28 GiB17.00 GiB138.97 GiB5.03 GiB7±37%
GLM-4.6MoEUD-IQ2_M357B113.56 GiB24.44 GiB138.59 GiB5.41 GiB10±37%
MiniMax-M2.5MoEQ4_K_S229B121.10 GiB16.47 GiB138.11 GiB5.89 GiB14±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ4_K_S229B121.10 GiB16.47 GiB138.11 GiB5.89 GiB14±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ4_K236B132.67 GiB4.48 GiB137.74 GiB6.26 GiB21±37%
DeepSeek-V2.5MoEQ4_K236B132.67 GiB4.48 GiB137.74 GiB6.26 GiB21±37%
DeepSeek-Coder-V2-InstructMoEQ4_K236B132.67 GiB4.48 GiB137.74 GiB6.26 GiB21±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ4_K_S236B124.51 GiB12.48 GiB137.57 GiB6.43 GiB14±37%
Qwen3-VL-235B-A22B-InstructMoEQ4_K_S236B124.51 GiB12.48 GiB137.57 GiB6.43 GiB14±37%
Qwen3-235B-A22BMoEQ4_K_S235B124.51 GiB12.48 GiB137.57 GiB6.43 GiB14±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q4_K_S235B124.51 GiB12.48 GiB137.57 GiB6.43 GiB14±37%
WizardLM-Uncensored-SuperCOT-StoryTelling-30bQ8_032.5B32.19 GiB103.59 GiB136.41 GiB7.59 GiB5±8.3%
archangel_sft-kto_llama30bQ8_032.5B32.19 GiB103.59 GiB136.41 GiB7.59 GiB5±8.3%
Wizard-Vicuna-30B-UncensoredQ8_032.5B32.19 GiB103.59 GiB136.41 GiB7.59 GiB5±8.3%
Nex-N2-ProMoEIQ2_M397B132.87 GiB1.99 GiB135.47 GiB8.53 GiB27±37%
ERNIE-4.5-300B-A47B-PTQ3_K_S300B120.25 GiB14.34 GiB135.27 GiB8.73 GiB5±8.3%
Trinity-Large-ThinkingMoEQ2_K_L399B130.22 GiB4.40 GiB135.20 GiB8.80 GiB24±37%
Hermes-4-405BIQ2_XXS406B99.91 GiB33.47 GiB134.21 GiB9.79 GiB5±8.3%
Hermes-3-Llama-3.1-405BIQ2_XXS406B99.91 GiB33.47 GiB134.21 GiB9.79 GiB5±8.3%
MiniMax-M3MoEUD-IQ2_M427B124.99 GiB7.97 GiB133.52 GiB10.48 GiB19±37%
GLM-4.6-REAP-268B-A32BMoEQ3_K_S269B108.47 GiB24.44 GiB133.50 GiB10.50 GiB10±37%
command-a-plus-05-2026-bf16MoEQ4_1219B129.29 GiB2.35 GiB132.20 GiB11.80 GiB20±37%
Qwen3.6-35B-A3B-abliterated-MAXMoEF3235.1B129.13 GiB1.33 GiB131.01 GiB12.99 GiB26±37%
DeepSeek-V4-FlashMoEUD-IQ4_NL291B128.43 GiB0.03 GiB129.06 GiB14.94 GiB31±37%
DeepSeek-V4-Flash-0731MoEUD-IQ4_NL304B127.28 GiB0.03 GiB127.91 GiB16.09 GiB31±37%
granite-34b-code-base-8kF3233.7B125.60 GiB0.00 GiB126.23 GiB17.77 GiB5±8.3%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ8_0124B119.65 GiB5.84 GiB126.04 GiB17.96 GiB20±37%
Qwen3.5-122B-A10BMoEQ8_0125B123.49 GiB1.59 GiB125.66 GiB18.34 GiB27±37%
Mistral-Medium-3.5-128BQ6_K_L128B101.13 GiB23.38 GiB125.21 GiB18.79 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 M2 Ultra run?
2092 of 2118 indexed open-weight models fit a Apple M2 Ultra at 131,072 context with q8_0 KV cache, the largest being dots.llm1.inst at Q4_0. That covers text, vision-language, image, video and speech models.
How much usable memory does a Apple M2 Ultra actually have?
Its nameplate is 192 GB, but about 133.92 GiB is available to a model once driver and compositor overhead is accounted for, and only 144 GB of the pool can be allocated to the GPU at all.
Is a Apple M2 Ultra fast for local AI?
Its memory bandwidth is 819 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.