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 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 64K context

largest quantization that fits, per model · 2100 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Mixtral-8x22B-Instruct-v0.1MoEQ8_0141B139.16 GiB3.94 GiB143.71 GiB0.29 GiB9±37%
Mixtral-8x22B-v0.1MoEQ8_0141B139.16 GiB3.94 GiB143.71 GiB0.29 GiB9±37%
Mixtral-8x22B-v0.1MoEQ8_0141B139.15 GiB3.94 GiB143.70 GiB0.30 GiB9±37%
command-a-plus-05-2026-bf16MoEQ5_K_S219B141.69 GiB0.68 GiB142.93 GiB1.07 GiB20±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEQ8_0139B137.78 GiB4.36 GiB142.68 GiB1.32 GiB18±37%
GLM-4.5MoEUD-IQ3_XXS358B135.21 GiB6.47 GiB142.27 GiB1.73 GiB17±37%
GLM-4.7MoEUD-IQ3_XXS358B135.15 GiB6.47 GiB142.21 GiB1.79 GiB17±37%
MiMo-V2.5MoEKV unresolvedUD-IQ4_XS311B139.18 GiB2.11 GiB141.89 GiB2.11 GiB24±37%
GLM-4.6MoEUD-IQ3_XXS357B134.76 GiB6.47 GiB141.82 GiB2.18 GiB17±37%
Qwen2.5-72BF1672.7B135.44 GiB5.63 GiB141.74 GiB2.26 GiB5±8.3%
Kimi-Dev-72BBF1672.7B135.44 GiB5.63 GiB141.74 GiB2.26 GiB5±8.3%
Qwen2.5-VL-72B-InstructBF1673.4B135.44 GiB5.63 GiB141.74 GiB2.26 GiB5±8.3%
Llama-3_1-Nemotron-51B-InstructF1651.5B95.94 GiB45.00 GiB141.63 GiB2.37 GiB5±8.3%
Qwen3.5-REAP-212B-A17BMoEQ5_K_M212B139.97 GiB0.53 GiB141.10 GiB2.90 GiB25±37%
Qwen3-235B-A22B-Instruct-2507MoEQ4_1235B137.20 GiB3.30 GiB141.09 GiB2.91 GiB18±37%
Qwen3-235B-A22B-Thinking-2507MoEQ4_1235B137.20 GiB3.30 GiB141.09 GiB2.91 GiB18±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ4_1236B137.12 GiB3.30 GiB141.01 GiB2.99 GiB18±37%
Qwen3-VL-235B-A22B-InstructMoEQ4_1236B137.12 GiB3.30 GiB141.01 GiB2.99 GiB18±37%
Qwen3-235B-A22BMoEQ4_1235B137.12 GiB3.30 GiB141.01 GiB2.99 GiB18±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q4_1235B137.12 GiB3.30 GiB141.01 GiB2.99 GiB18±37%
GLM-4.6-REAP-268B-A32BMoEIQ4_XS269B133.87 GiB6.47 GiB140.93 GiB3.07 GiB16±37%
Qwen3.5-397B-A17BMoEUD-IQ3_XXS403B139.55 GiB0.53 GiB140.68 GiB3.32 GiB29±37%
DeepSeek-V3.1-TerminusMoEIQ1_M685B138.82 GiB1.21 GiB140.66 GiB3.34 GiB25±37%
DeepSeek-V3.2MoEIQ1_M685B138.82 GiB1.21 GiB140.66 GiB3.34 GiB25±37%
cogito-671b-v2.1MoEIQ1_M671B138.82 GiB1.21 GiB140.65 GiB3.35 GiB25±37%
DeepSeek-V3-0324MoEIQ1_M685B138.66 GiB1.21 GiB140.49 GiB3.51 GiB25±37%
r1-1776MoEIQ1_M671B138.66 GiB1.21 GiB140.49 GiB3.51 GiB25±37%
DeepSeek-R1MoEIQ1_M685B138.66 GiB1.21 GiB140.49 GiB3.51 GiB25±37%
Step-3.7-FlashQ5_K_L201B132.58 GiB7.04 GiB140.20 GiB3.80 GiB5±8.3%
grok-2MoEIQ4_XS270B134.85 GiB4.50 GiB140.04 GiB3.96 GiB9±37%
Hy3MoEIQ3_M299B133.80 GiB5.63 GiB140.01 GiB3.99 GiB19±37%
MiMo-V2-FlashMoEKV unresolvedQ3_K_M310B137.19 GiB2.11 GiB139.90 GiB4.10 GiB24±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ2_K_L402B135.87 GiB3.38 GiB139.82 GiB4.18 GiB27±37%
GLM-4.6-Derestricted-v3MoEIQ3_XXS357B132.66 GiB6.47 GiB139.72 GiB4.28 GiB17±37%
Qwen3.5-REAP-262B-A17BMoEQ4_K_S262B138.50 GiB0.53 GiB139.63 GiB4.37 GiB26±37%
DeepSeek-R1-0528MoEIQ1_M685B137.32 GiB1.21 GiB139.16 GiB4.84 GiB25±37%
DeepSeek-V3.1MoEIQ1_M685B137.32 GiB1.21 GiB139.16 GiB4.84 GiB25±37%
Ornith-1.0-397BMoEUD-IQ3_XXS397B137.46 GiB0.53 GiB138.59 GiB5.41 GiB29±37%
Llama-3_3-Nemotron-Super-49B-v1_5BF1649.9B92.89 GiB45.00 GiB138.59 GiB5.41 GiB5±8.3%
Valkyrie-49B-v2.1BF1649.9B92.89 GiB45.00 GiB138.59 GiB5.41 GiB5±8.3%
Llama-3_3-Nemotron-Super-49B-v1BF1649.9B92.89 GiB45.00 GiB138.59 GiB5.41 GiB5±8.3%
MiniMax-M2.7MoEQ4_1229B133.65 GiB4.36 GiB138.55 GiB5.45 GiB21±37%
step-3.5-flashQ5_K_L199B130.88 GiB7.04 GiB138.50 GiB5.50 GiB5±8.3%
MiniMax-M2.1MoEQ4_1229B133.47 GiB4.36 GiB138.36 GiB5.64 GiB21±37%
MiniMax-M2MoEQ4_1229B133.47 GiB4.36 GiB138.36 GiB5.64 GiB21±37%
MiniMax-M2.5MoEQ4_1229B133.39 GiB4.36 GiB138.29 GiB5.71 GiB22±37%
MiniMax-M3MoEIQ2_M427B135.58 GiB2.11 GiB138.25 GiB5.75 GiB24±37%
dots.llm1.instMoEQ6_K143B120.06 GiB17.44 GiB138.07 GiB5.93 GiB12±37%
Apertus-70B-Instruct-2509BF1670.6B131.51 GiB5.63 GiB137.87 GiB6.13 GiB5±8.3%
Llama-3.3-70B-InstructF1670.6B131.43 GiB5.63 GiB137.73 GiB6.27 GiB5±8.3%
Hermes-4-70BBF1670.6B131.43 GiB5.63 GiB137.73 GiB6.27 GiB5±8.3%
Llama-3.1-70BF1670.6B131.43 GiB5.63 GiB137.73 GiB6.27 GiB5±8.3%
DeepSeek-R1-Distill-Llama-70BF1670.6B131.43 GiB5.63 GiB137.73 GiB6.27 GiB5±8.3%
Athene-70BBF1670.6B131.43 GiB5.63 GiB137.73 GiB6.27 GiB5±8.3%
Hermes-3-Llama-3.1-70BBF1670.6B131.43 GiB5.63 GiB137.73 GiB6.27 GiB5±8.3%
Meta-Llama-3-70B-Instruct-abliterated-v3.5BF1670.6B131.43 GiB5.63 GiB137.73 GiB6.27 GiB5±8.3%
L3.3-70B-Magnum-DiamondBF1670.6B131.43 GiB5.63 GiB137.73 GiB6.27 GiB5±8.3%
ERNIE-4.5-300B-A47B-PTQ3_K_M300B132.78 GiB3.80 GiB137.25 GiB6.75 GiB5±8.3%
Trinity-Large-PreviewMoEQ2_K_L399B135.17 GiB1.28 GiB137.02 GiB6.98 GiB29±37%
Hermes-3-Llama-3.1-405BIQ2_M406B127.28 GiB8.86 GiB136.97 GiB7.03 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 65,536 context with q4_0 KV cache, the largest being Mixtral-8x22B-Instruct-v0.1 at Q8_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.