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. 2100 of 2118 indexed models fit at 8K context with q4_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 1805vision language 191image 2audio tts 21audio asr 39video 16embedding 26

What fits at 8K context

largest quantization that fits, per model · 2100 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Trinity-Large-PreviewMoEUD-IQ2_XXS399B142.94 GiB0.35 GiB143.87 GiB0.13 GiB30±37%
GLM-4.6-REAP-268B-A32BMoEQ4_K_S269B142.42 GiB0.81 GiB143.82 GiB0.18 GiB19±37%
grok-2MoEIQ4_NL270B142.48 GiB0.56 GiB143.73 GiB0.27 GiB9±37%
Trinity-Large-TrueBaseMoEI1-IQ3_XXS399B142.48 GiB0.35 GiB143.41 GiB0.59 GiB30±37%
MiniMax-M3MoEQ2_K_L427B142.57 GiB0.26 GiB143.40 GiB0.60 GiB26±37%
MiMo-V2.5MoEKV unresolvedUD-IQ4_NL311B142.12 GiB0.26 GiB142.98 GiB1.02 GiB26±37%
Hermes-3-Llama-3.1-405BQ2_K_L406B140.98 GiB1.11 GiB142.92 GiB1.08 GiB5±8.3%
command-a-plus-05-2026-bf16MoEQ5_K_S219B141.69 GiB0.19 GiB142.43 GiB1.57 GiB21±37%
Solar-Open2-250BMoEQ4_K_M250B141.39 GiB0.42 GiB142.39 GiB1.61 GiB27±37%
GLM-4.7-REAP-218B-A32BMoEQ5_K_S218B140.74 GiB0.81 GiB142.14 GiB1.86 GiB18±37%
Hermes-4-405BQ2_K406B139.07 GiB1.11 GiB141.01 GiB2.99 GiB5±8.3%
Qwen3.5-REAP-212B-A17BMoEQ5_K_M212B139.97 GiB0.07 GiB140.63 GiB3.37 GiB25±37%
Qwen3-Coder-480B-A35B-InstructMoEUD-IQ1_M480B139.43 GiB0.54 GiB140.56 GiB3.44 GiB22±37%
Mixtral-8x22B-Instruct-v0.1MoEQ8_0141B139.16 GiB0.49 GiB140.27 GiB3.73 GiB9±37%
Mixtral-8x22B-v0.1MoEQ8_0141B139.16 GiB0.49 GiB140.26 GiB3.74 GiB9±37%
Mixtral-8x22B-v0.1MoEQ8_0141B139.15 GiB0.49 GiB140.26 GiB3.74 GiB9±37%
Qwen3.5-397B-A17BMoEUD-IQ3_XXS403B139.55 GiB0.07 GiB140.22 GiB3.78 GiB30±37%
GLM-4.5MoEIQ3_XS358B138.63 GiB0.81 GiB140.03 GiB3.97 GiB22±37%
GLM-4.7MoEIQ3_XS358B138.48 GiB0.81 GiB139.87 GiB4.13 GiB22±37%
DeepSeek-V3.1-TerminusMoEIQ1_M685B138.82 GiB0.15 GiB139.60 GiB4.40 GiB27±37%
DeepSeek-V3.2MoEIQ1_M685B138.82 GiB0.15 GiB139.60 GiB4.40 GiB27±37%
cogito-671b-v2.1MoEIQ1_M671B138.82 GiB0.15 GiB139.60 GiB4.40 GiB27±37%
DeepSeek-V3-0324MoEIQ1_M685B138.66 GiB0.15 GiB139.44 GiB4.56 GiB27±37%
r1-1776MoEIQ1_M671B138.66 GiB0.15 GiB139.44 GiB4.56 GiB27±37%
DeepSeek-R1MoEIQ1_M685B138.66 GiB0.15 GiB139.44 GiB4.56 GiB27±37%
GLM-4.6-Derestricted-v3MoEIQ3_XS357B138.01 GiB0.81 GiB139.41 GiB4.59 GiB22±37%
GLM-4.6MoEIQ3_XS357B138.01 GiB0.81 GiB139.41 GiB4.59 GiB22±37%
Qwen3.5-REAP-262B-A17BMoEQ4_K_S262B138.50 GiB0.07 GiB139.17 GiB4.83 GiB27±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEQ8_0139B137.78 GiB0.54 GiB138.86 GiB5.14 GiB22±37%
Qwen3-235B-A22B-Instruct-2507MoEQ4_1235B137.20 GiB0.41 GiB138.19 GiB5.81 GiB21±37%
Qwen3-235B-A22B-Thinking-2507MoEQ4_1235B137.20 GiB0.41 GiB138.19 GiB5.81 GiB21±37%
Step-3.7-FlashUD-Q5_K_M201B136.43 GiB1.13 GiB138.15 GiB5.85 GiB5±8.3%
Ornith-1.0-397BMoEUD-IQ3_XXS397B137.46 GiB0.07 GiB138.13 GiB5.87 GiB30±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ4_1236B137.12 GiB0.41 GiB138.11 GiB5.89 GiB21±37%
Qwen3-VL-235B-A22B-InstructMoEQ4_1236B137.12 GiB0.41 GiB138.11 GiB5.89 GiB21±37%
Qwen3-235B-A22BMoEQ4_1235B137.12 GiB0.41 GiB138.11 GiB5.89 GiB21±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q4_1235B137.12 GiB0.41 GiB138.11 GiB5.89 GiB21±37%
DeepSeek-R1-0528MoEIQ1_M685B137.32 GiB0.15 GiB138.10 GiB5.90 GiB27±37%
DeepSeek-V3.1MoEIQ1_M685B137.32 GiB0.15 GiB138.10 GiB5.90 GiB27±37%
MiMo-V2-FlashMoEKV unresolvedQ3_K_M310B137.19 GiB0.26 GiB138.05 GiB5.95 GiB27±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ2_K_L402B135.87 GiB0.42 GiB136.87 GiB7.13 GiB33±37%
Qwen2.5-72BF1672.7B135.44 GiB0.70 GiB136.82 GiB7.18 GiB5±8.3%
Kimi-Dev-72BBF1672.7B135.44 GiB0.70 GiB136.82 GiB7.18 GiB5±8.3%
Qwen2.5-VL-72B-InstructBF1673.4B135.44 GiB0.70 GiB136.82 GiB7.18 GiB5±8.3%
Hy3MoEIQ3_M299B133.80 GiB0.70 GiB135.09 GiB8.91 GiB25±37%
MiniMax-M2.7MoEQ4_1229B133.65 GiB0.54 GiB134.73 GiB9.27 GiB27±37%
MiniMax-M2.1MoEQ4_1229B133.47 GiB0.54 GiB134.55 GiB9.45 GiB27±37%
MiniMax-M2MoEQ4_1229B133.47 GiB0.54 GiB134.55 GiB9.45 GiB27±37%
MiniMax-M2.5MoEQ4_1229B133.39 GiB0.54 GiB134.47 GiB9.53 GiB27±37%
ERNIE-4.5-300B-A47B-PTQ3_K_M300B132.78 GiB0.47 GiB133.93 GiB10.07 GiB5±8.3%
Nex-N2-ProMoEIQ2_M397B132.87 GiB0.07 GiB133.54 GiB10.46 GiB31±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ4_K236B132.67 GiB0.15 GiB133.41 GiB10.59 GiB26±37%
DeepSeek-V2.5MoEQ4_K236B132.67 GiB0.15 GiB133.41 GiB10.59 GiB26±37%
DeepSeek-Coder-V2-InstructMoEQ4_K236B132.67 GiB0.15 GiB133.41 GiB10.59 GiB26±37%
Apertus-70B-Instruct-2509BF1670.6B131.51 GiB0.70 GiB132.95 GiB11.05 GiB5±8.3%
Llama-3.3-70B-InstructF1670.6B131.43 GiB0.70 GiB132.81 GiB11.19 GiB5±8.3%
Hermes-4-70BBF1670.6B131.43 GiB0.70 GiB132.81 GiB11.19 GiB5±8.3%
Llama-3.1-70BF1670.6B131.43 GiB0.70 GiB132.81 GiB11.19 GiB5±8.3%
DeepSeek-R1-Distill-Llama-70BF1670.6B131.43 GiB0.70 GiB132.81 GiB11.19 GiB5±8.3%
Athene-70BBF1670.6B131.43 GiB0.70 GiB132.81 GiB11.19 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 M2 Ultra run?
2100 of 2118 indexed open-weight models fit a Apple M2 Ultra at 8,192 context with q4_0 KV cache, the largest being Trinity-Large-Preview at UD-IQ2_XXS. 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.
Apple M2 Ultra — what AI models can it run locally? — ossmodeldb