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 16K 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 16K context

largest quantization that fits, per model · 2100 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
GLM-4.6-REAP-268B-A32BMoEIQ4_NL269B141.57 GiB1.62 GiB143.78 GiB0.22 GiB19±37%
MiniMax-M3MoEQ2_K_L427B142.57 GiB0.53 GiB143.67 GiB0.33 GiB26±37%
Trinity-Large-TrueBaseMoEI1-IQ3_XXS399B142.48 GiB0.49 GiB143.54 GiB0.46 GiB30±37%
MiMo-V2.5MoEKV unresolvedUD-IQ4_NL311B142.12 GiB0.53 GiB143.24 GiB0.76 GiB26±37%
GLM-4.7-REAP-218B-A32BMoEQ5_K_S218B140.74 GiB1.62 GiB142.95 GiB1.05 GiB17±37%
Solar-Open2-250BMoEQ4_K_M250B141.39 GiB0.84 GiB142.81 GiB1.19 GiB27±37%
command-a-plus-05-2026-bf16MoEQ5_K_S219B141.69 GiB0.26 GiB142.50 GiB1.50 GiB21±37%
Hermes-3-Llama-3.1-405BQ2_K406B139.07 GiB2.21 GiB142.11 GiB1.89 GiB5±8.3%
Hermes-4-405BQ2_K406B139.07 GiB2.21 GiB142.11 GiB1.89 GiB5±8.3%
Qwen3-Coder-480B-A35B-InstructMoEUD-IQ1_M480B139.43 GiB1.09 GiB141.11 GiB2.89 GiB22±37%
GLM-4.5MoEIQ3_XS358B138.63 GiB1.62 GiB140.84 GiB3.16 GiB21±37%
Mixtral-8x22B-Instruct-v0.1MoEQ8_0141B139.16 GiB0.98 GiB140.76 GiB3.24 GiB9±37%
Mixtral-8x22B-v0.1MoEQ8_0141B139.16 GiB0.98 GiB140.75 GiB3.25 GiB9±37%
Mixtral-8x22B-v0.1MoEQ8_0141B139.15 GiB0.98 GiB140.75 GiB3.25 GiB9±37%
Qwen3.5-REAP-212B-A17BMoEQ5_K_M212B139.97 GiB0.13 GiB140.70 GiB3.30 GiB25±37%
GLM-4.7MoEIQ3_XS358B138.48 GiB1.62 GiB140.68 GiB3.32 GiB21±37%
Qwen3.5-397B-A17BMoEUD-IQ3_XXS403B139.55 GiB0.13 GiB140.28 GiB3.72 GiB30±37%
GLM-4.6-Derestricted-v3MoEIQ3_XS357B138.01 GiB1.62 GiB140.22 GiB3.78 GiB21±37%
GLM-4.6MoEIQ3_XS357B138.01 GiB1.62 GiB140.22 GiB3.78 GiB21±37%
DeepSeek-V3.1-TerminusMoEIQ1_M685B138.82 GiB0.30 GiB139.75 GiB4.25 GiB27±37%
DeepSeek-V3.2MoEIQ1_M685B138.82 GiB0.30 GiB139.75 GiB4.25 GiB27±37%
cogito-671b-v2.1MoEIQ1_M671B138.82 GiB0.30 GiB139.75 GiB4.25 GiB27±37%
DeepSeek-V3-0324MoEIQ1_M685B138.66 GiB0.30 GiB139.59 GiB4.41 GiB27±37%
r1-1776MoEIQ1_M671B138.66 GiB0.30 GiB139.59 GiB4.41 GiB27±37%
DeepSeek-R1MoEIQ1_M685B138.66 GiB0.30 GiB139.59 GiB4.41 GiB27±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEQ8_0139B137.78 GiB1.09 GiB139.41 GiB4.59 GiB22±37%
Qwen3.5-REAP-262B-A17BMoEQ4_K_S262B138.50 GiB0.13 GiB139.23 GiB4.77 GiB27±37%
Step-3.7-FlashUD-Q5_K_M201B136.43 GiB1.98 GiB138.99 GiB5.01 GiB5±8.3%
Qwen3-235B-A22B-Instruct-2507MoEQ4_1235B137.20 GiB0.83 GiB138.61 GiB5.39 GiB21±37%
Qwen3-235B-A22B-Thinking-2507MoEQ4_1235B137.20 GiB0.83 GiB138.61 GiB5.39 GiB21±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ4_1236B137.12 GiB0.83 GiB138.53 GiB5.47 GiB21±37%
Qwen3-VL-235B-A22B-InstructMoEQ4_1236B137.12 GiB0.83 GiB138.53 GiB5.47 GiB21±37%
Qwen3-235B-A22BMoEQ4_1235B137.12 GiB0.83 GiB138.53 GiB5.47 GiB21±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q4_1235B137.12 GiB0.83 GiB138.53 GiB5.47 GiB21±37%
MiMo-V2-FlashMoEKV unresolvedQ3_K_M310B137.19 GiB0.53 GiB138.31 GiB5.69 GiB26±37%
DeepSeek-R1-0528MoEIQ1_M685B137.32 GiB0.30 GiB138.25 GiB5.75 GiB27±37%
DeepSeek-V3.1MoEIQ1_M685B137.32 GiB0.30 GiB138.25 GiB5.75 GiB27±37%
Ornith-1.0-397BMoEUD-IQ3_XXS397B137.46 GiB0.13 GiB138.19 GiB5.81 GiB30±37%
Qwen2.5-72BF1672.7B135.44 GiB1.41 GiB137.52 GiB6.48 GiB5±8.3%
Kimi-Dev-72BBF1672.7B135.44 GiB1.41 GiB137.52 GiB6.48 GiB5±8.3%
Qwen2.5-VL-72B-InstructBF1673.4B135.44 GiB1.41 GiB137.52 GiB6.48 GiB5±8.3%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ2_K_L402B135.87 GiB0.84 GiB137.29 GiB6.71 GiB32±37%
grok-2MoEIQ4_XS270B134.85 GiB1.13 GiB136.67 GiB7.33 GiB9±37%
Trinity-Large-PreviewMoEQ2_K_L399B135.17 GiB0.49 GiB136.23 GiB7.77 GiB31±37%
Hy3MoEIQ3_M299B133.80 GiB1.41 GiB135.79 GiB8.21 GiB24±37%
MiniMax-M2.7MoEQ4_1229B133.65 GiB1.09 GiB135.28 GiB8.72 GiB26±37%
MiniMax-M2.1MoEQ4_1229B133.47 GiB1.09 GiB135.09 GiB8.91 GiB26±37%
MiniMax-M2MoEQ4_1229B133.47 GiB1.09 GiB135.09 GiB8.91 GiB26±37%
MiniMax-M2.5MoEQ4_1229B133.39 GiB1.09 GiB135.02 GiB8.98 GiB26±37%
ERNIE-4.5-300B-A47B-PTQ3_K_M300B132.78 GiB0.95 GiB134.40 GiB9.60 GiB5±8.3%
Apertus-70B-Instruct-2509BF1670.6B131.51 GiB1.41 GiB133.65 GiB10.35 GiB5±8.3%
Nex-N2-ProMoEIQ2_M397B132.87 GiB0.13 GiB133.60 GiB10.40 GiB31±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ4_K236B132.67 GiB0.30 GiB133.56 GiB10.44 GiB26±37%
DeepSeek-V2.5MoEQ4_K236B132.67 GiB0.30 GiB133.56 GiB10.44 GiB26±37%
DeepSeek-Coder-V2-InstructMoEQ4_K236B132.67 GiB0.30 GiB133.56 GiB10.44 GiB26±37%
Llama-3.3-70B-InstructF1670.6B131.43 GiB1.41 GiB133.51 GiB10.49 GiB5±8.3%
Hermes-4-70BBF1670.6B131.43 GiB1.41 GiB133.51 GiB10.49 GiB5±8.3%
Llama-3.1-70BF1670.6B131.43 GiB1.41 GiB133.51 GiB10.49 GiB5±8.3%
DeepSeek-R1-Distill-Llama-70BF1670.6B131.43 GiB1.41 GiB133.51 GiB10.49 GiB5±8.3%
Athene-70BBF1670.6B131.43 GiB1.41 GiB133.51 GiB10.49 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 16,384 context with q4_0 KV cache, the largest being GLM-4.6-REAP-268B-A32B at IQ4_NL. 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.