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 64K 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 1805vision language 191image 2audio tts 21audio asr 39video 16embedding 26

What fits at 64K context

largest quantization that fits, per model · 2100 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Qwen3-VL-235B-A22B-ThinkingMoEQ4_1236B137.12 GiB6.24 GiB143.94 GiB0.06 GiB16±37%
Qwen3-VL-235B-A22B-InstructMoEQ4_1236B137.12 GiB6.24 GiB143.94 GiB0.06 GiB16±37%
Qwen3-235B-A22BMoEQ4_1235B137.12 GiB6.24 GiB143.94 GiB0.06 GiB16±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q4_1235B137.12 GiB6.24 GiB143.94 GiB0.06 GiB16±37%
MiMo-V2.5MoEKV unresolvedUD-IQ4_XS311B139.18 GiB3.98 GiB143.76 GiB0.24 GiB21±37%
command-a-plus-05-2026-bf16MoEQ5_K_S219B141.69 GiB1.29 GiB143.53 GiB0.47 GiB20±37%
Apertus-70B-Instruct-2509BF1670.6B131.51 GiB10.63 GiB142.87 GiB1.13 GiB5±8.3%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ2_K_L402B135.87 GiB6.38 GiB142.82 GiB1.18 GiB22±37%
Llama-3.3-70B-InstructF1670.6B131.43 GiB10.63 GiB142.73 GiB1.27 GiB5±8.3%
Hermes-4-70BBF1670.6B131.43 GiB10.63 GiB142.73 GiB1.27 GiB5±8.3%
Llama-3.1-70BF1670.6B131.43 GiB10.63 GiB142.73 GiB1.27 GiB5±8.3%
DeepSeek-R1-Distill-Llama-70BF1670.6B131.43 GiB10.63 GiB142.73 GiB1.27 GiB5±8.3%
Athene-70BBF1670.6B131.43 GiB10.63 GiB142.73 GiB1.27 GiB5±8.3%
Hermes-3-Llama-3.1-70BBF1670.6B131.43 GiB10.63 GiB142.73 GiB1.27 GiB5±8.3%
Meta-Llama-3-70B-Instruct-abliterated-v3.5BF1670.6B131.43 GiB10.63 GiB142.73 GiB1.27 GiB5±8.3%
L3.3-70B-Magnum-DiamondBF1670.6B131.43 GiB10.63 GiB142.73 GiB1.27 GiB5±8.3%
Step-3.7-FlashUD-Q5_K_S201B128.59 GiB13.30 GiB142.47 GiB1.53 GiB5±8.3%
MiniMax-M2.7MoEQ4_1229B133.65 GiB8.23 GiB142.42 GiB1.58 GiB18±37%
MiniMax-M2.1MoEQ4_1229B133.47 GiB8.23 GiB142.24 GiB1.76 GiB18±37%
MiniMax-M2MoEQ4_1229B133.47 GiB8.23 GiB142.24 GiB1.76 GiB18±37%
MiniMax-M2.5MoEQ4_1229B133.39 GiB8.23 GiB142.16 GiB1.84 GiB18±37%
MiMo-V2-FlashMoEKV unresolvedQ3_K_M310B137.19 GiB3.98 GiB141.77 GiB2.23 GiB22±37%
DeepSeek-V3.1-TerminusMoEIQ1_M685B138.82 GiB2.28 GiB141.73 GiB2.27 GiB24±37%
DeepSeek-V3.2MoEIQ1_M685B138.82 GiB2.28 GiB141.73 GiB2.27 GiB24±37%
cogito-671b-v2.1MoEIQ1_M671B138.82 GiB2.28 GiB141.73 GiB2.27 GiB24±37%
DeepSeek-V3-0324MoEIQ1_M685B138.66 GiB2.28 GiB141.57 GiB2.43 GiB24±37%
r1-1776MoEIQ1_M671B138.66 GiB2.28 GiB141.57 GiB2.43 GiB24±37%
DeepSeek-R1MoEIQ1_M685B138.66 GiB2.28 GiB141.57 GiB2.43 GiB24±37%
Qwen3.5-REAP-212B-A17BMoEQ5_K_M212B139.97 GiB1.00 GiB141.56 GiB2.44 GiB24±37%
Qwen3.5-397B-A17BMoEUD-IQ3_XXS403B139.55 GiB1.00 GiB141.15 GiB2.85 GiB28±37%
GLM-4.7-REAP-218B-A32BMoEQ4_1218B128.16 GiB12.22 GiB140.97 GiB3.03 GiB12±37%
ERNIE-4.5-300B-A47B-PTQ3_K_M300B132.78 GiB7.17 GiB140.62 GiB3.38 GiB5±8.3%
step-3.5-flashQ5_K_S199B126.55 GiB13.30 GiB140.43 GiB3.57 GiB5±8.3%
DeepSeek-R1-0528MoEIQ1_M685B137.32 GiB2.28 GiB140.23 GiB3.77 GiB24±37%
DeepSeek-V3.1MoEIQ1_M685B137.32 GiB2.28 GiB140.23 GiB3.77 GiB24±37%
MiniMax-M3MoEIQ2_M427B135.58 GiB3.98 GiB140.13 GiB3.87 GiB22±37%
Qwen3-235B-A22B-Instruct-2507MoEQ4_K_L235B133.28 GiB6.24 GiB140.11 GiB3.89 GiB16±37%
Qwen3.5-REAP-262B-A17BMoEQ4_K_S262B138.50 GiB1.00 GiB140.10 GiB3.90 GiB26±37%
Qwen3-235B-A22B-Thinking-2507MoEQ4_K_M235B132.85 GiB6.24 GiB139.68 GiB4.32 GiB16±37%
Hy3MoEIQ3_XS299B128.00 GiB10.63 GiB139.21 GiB4.79 GiB16±37%
Ornith-1.0-397BMoEUD-IQ3_XXS397B137.46 GiB1.00 GiB139.06 GiB4.94 GiB28±37%
Trinity-Large-PreviewMoEQ2_K_L399B135.17 GiB2.41 GiB138.16 GiB5.84 GiB27±37%
Trinity-Large-TrueBaseMoEI1-Q2_K399B134.81 GiB2.41 GiB137.80 GiB6.20 GiB27±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ4_K_M229B128.84 GiB8.23 GiB137.61 GiB6.39 GiB18±37%
Yi-1.5-34BF3234.4B128.11 GiB7.97 GiB136.71 GiB7.29 GiB5±8.3%
Llama-3_1-Nemotron-51B-InstructQ8_051.5B50.97 GiB85.00 GiB136.66 GiB7.34 GiB5±8.3%
Devstral-2-123B-Instruct-2512Q8_0125B123.73 GiB11.69 GiB136.13 GiB7.87 GiB5±8.3%
Mistral-Medium-3.5-128BQ8_0128B123.73 GiB11.69 GiB136.12 GiB7.88 GiB5±8.3%
DeepSeek-Coder-V2-Instruct-0724MoEQ4_K236B132.67 GiB2.24 GiB135.50 GiB8.50 GiB23±37%
DeepSeek-V2.5MoEQ4_K236B132.67 GiB2.24 GiB135.50 GiB8.50 GiB23±37%
DeepSeek-Coder-V2-InstructMoEQ4_K236B132.67 GiB2.24 GiB135.50 GiB8.50 GiB23±37%
dots.llm1.instMoEQ5_K_M143B101.84 GiB32.94 GiB135.36 GiB8.64 GiB9±37%
GLM-4.5MoEQ2_K_L358B122.32 GiB12.22 GiB135.13 GiB8.87 GiB15±37%
GLM-4.7MoEQ2_K_L358B122.32 GiB12.22 GiB135.13 GiB8.87 GiB15±37%
Llama-3_3-Nemotron-Super-49B-v1_5Q8_049.9B49.36 GiB85.00 GiB135.05 GiB8.95 GiB5±8.3%
Valkyrie-49B-v2.1Q8_049.9B49.36 GiB85.00 GiB135.05 GiB8.95 GiB5±8.3%
Llama-3_3-Nemotron-Super-49B-v1Q8_049.9B49.36 GiB85.00 GiB135.05 GiB8.95 GiB5±8.3%
GLM-4.6MoEQ2_K_L357B121.85 GiB12.22 GiB134.66 GiB9.34 GiB15±37%
Hermes-4-405BIQ2_S406B117.02 GiB16.73 GiB134.58 GiB9.42 GiB5±8.3%
Nex-N2-ProMoEIQ2_M397B132.87 GiB1.00 GiB134.47 GiB9.53 GiB29±37%
●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 65,536 context with q8_0 KV cache, the largest being Qwen3-VL-235B-A22B-Thinking at Q4_1. 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