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 32K 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
vision language 191text 1805image 2audio tts 21audio asr 39video 16embedding 26

What fits at 32K context

largest quantization that fits, per model · 2100 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Step-3.7-FlashUD-Q5_K_M201B136.43 GiB6.92 GiB143.94 GiB0.06 GiB5±8.3%
Mixtral-8x22B-Instruct-v0.1MoEQ8_0141B139.16 GiB3.72 GiB143.49 GiB0.51 GiB9±37%
Mixtral-8x22B-v0.1MoEQ8_0141B139.16 GiB3.72 GiB143.49 GiB0.51 GiB9±37%
Mixtral-8x22B-v0.1MoEQ8_0141B139.15 GiB3.72 GiB143.48 GiB0.52 GiB9±37%
command-a-plus-05-2026-bf16MoEQ5_K_S219B141.69 GiB0.76 GiB143.00 GiB1.00 GiB20±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEQ8_0139B137.78 GiB4.12 GiB142.44 GiB1.56 GiB19±37%
GLM-4.5MoEUD-IQ3_XXS358B135.21 GiB6.11 GiB141.91 GiB2.09 GiB17±37%
GLM-4.7MoEUD-IQ3_XXS358B135.15 GiB6.11 GiB141.85 GiB2.15 GiB18±37%
MiMo-V2.5MoEKV unresolvedUD-IQ4_XS311B139.18 GiB1.99 GiB141.77 GiB2.23 GiB24±37%
GLM-4.6MoEUD-IQ3_XXS357B134.76 GiB6.11 GiB141.46 GiB2.54 GiB18±37%
Qwen2.5-72BF1672.7B135.44 GiB5.31 GiB141.43 GiB2.57 GiB5±8.3%
Kimi-Dev-72BBF1672.7B135.44 GiB5.31 GiB141.43 GiB2.57 GiB5±8.3%
Qwen2.5-VL-72B-InstructBF1673.4B135.44 GiB5.31 GiB141.43 GiB2.57 GiB5±8.3%
Qwen3.5-REAP-212B-A17BMoEQ5_K_M212B139.97 GiB0.50 GiB141.07 GiB2.93 GiB25±37%
Qwen3-235B-A22B-Instruct-2507MoEQ4_1235B137.20 GiB3.12 GiB140.90 GiB3.10 GiB18±37%
Qwen3-235B-A22B-Thinking-2507MoEQ4_1235B137.20 GiB3.12 GiB140.90 GiB3.10 GiB18±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ4_1236B137.12 GiB3.12 GiB140.82 GiB3.18 GiB18±37%
Qwen3-VL-235B-A22B-InstructMoEQ4_1236B137.12 GiB3.12 GiB140.82 GiB3.18 GiB18±37%
Qwen3-235B-A22BMoEQ4_1235B137.12 GiB3.12 GiB140.82 GiB3.18 GiB18±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q4_1235B137.12 GiB3.12 GiB140.82 GiB3.18 GiB18±37%
Qwen3.5-397B-A17BMoEUD-IQ3_XXS403B139.55 GiB0.50 GiB140.65 GiB3.35 GiB29±37%
DeepSeek-V3.1-TerminusMoEIQ1_M685B138.82 GiB1.14 GiB140.59 GiB3.41 GiB25±37%
DeepSeek-V3.2MoEIQ1_M685B138.82 GiB1.14 GiB140.59 GiB3.41 GiB25±37%
cogito-671b-v2.1MoEIQ1_M671B138.82 GiB1.14 GiB140.59 GiB3.41 GiB25±37%
GLM-4.6-REAP-268B-A32BMoEIQ4_XS269B133.87 GiB6.11 GiB140.57 GiB3.43 GiB16±37%
DeepSeek-V3-0324MoEIQ1_M685B138.66 GiB1.14 GiB140.43 GiB3.57 GiB25±37%
r1-1776MoEIQ1_M671B138.66 GiB1.14 GiB140.43 GiB3.57 GiB25±37%
DeepSeek-R1MoEIQ1_M685B138.66 GiB1.14 GiB140.43 GiB3.57 GiB25±37%
grok-2MoEIQ4_XS270B134.85 GiB4.25 GiB139.79 GiB4.21 GiB9±37%
MiMo-V2-FlashMoEKV unresolvedQ3_K_M310B137.19 GiB1.99 GiB139.78 GiB4.22 GiB24±37%
Hy3MoEIQ3_M299B133.80 GiB5.31 GiB139.70 GiB4.30 GiB19±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ2_K_L402B135.87 GiB3.19 GiB139.63 GiB4.37 GiB27±37%
Qwen3.5-REAP-262B-A17BMoEQ4_K_S262B138.50 GiB0.50 GiB139.60 GiB4.40 GiB27±37%
GLM-4.6-Derestricted-v3MoEIQ3_XXS357B132.66 GiB6.11 GiB139.36 GiB4.64 GiB18±37%
Llama-3_1-Nemotron-51B-InstructF1651.5B95.94 GiB42.50 GiB139.13 GiB4.87 GiB5±8.3%
DeepSeek-R1-0528MoEIQ1_M685B137.32 GiB1.14 GiB139.09 GiB4.91 GiB25±37%
DeepSeek-V3.1MoEIQ1_M685B137.32 GiB1.14 GiB139.09 GiB4.91 GiB25±37%
Ornith-1.0-397BMoEUD-IQ3_XXS397B137.46 GiB0.50 GiB138.56 GiB5.44 GiB29±37%
step-3.5-flashQ5_K_L199B130.88 GiB6.92 GiB138.38 GiB5.62 GiB5±8.3%
MiniMax-M2.7MoEQ4_1229B133.65 GiB4.12 GiB138.30 GiB5.70 GiB22±37%
MiniMax-M3MoEIQ2_M427B135.58 GiB1.99 GiB138.14 GiB5.86 GiB24±37%
MiniMax-M2.1MoEQ4_1229B133.47 GiB4.12 GiB138.12 GiB5.88 GiB22±37%
MiniMax-M2MoEQ4_1229B133.47 GiB4.12 GiB138.12 GiB5.88 GiB22±37%
MiniMax-M2.5MoEQ4_1229B133.39 GiB4.12 GiB138.04 GiB5.96 GiB22±37%
Apertus-70B-Instruct-2509BF1670.6B131.51 GiB5.31 GiB137.56 GiB6.44 GiB5±8.3%
Llama-3.3-70B-InstructF1670.6B131.43 GiB5.31 GiB137.41 GiB6.59 GiB5±8.3%
Hermes-4-70BBF1670.6B131.43 GiB5.31 GiB137.41 GiB6.59 GiB5±8.3%
Llama-3.1-70BF1670.6B131.43 GiB5.31 GiB137.41 GiB6.59 GiB5±8.3%
DeepSeek-R1-Distill-Llama-70BF1670.6B131.43 GiB5.31 GiB137.41 GiB6.59 GiB5±8.3%
Athene-70BBF1670.6B131.43 GiB5.31 GiB137.41 GiB6.59 GiB5±8.3%
Hermes-3-Llama-3.1-70BBF1670.6B131.43 GiB5.31 GiB137.41 GiB6.59 GiB5±8.3%
Meta-Llama-3-70B-Instruct-abliterated-v3.5BF1670.6B131.43 GiB5.31 GiB137.41 GiB6.59 GiB5±8.3%
L3.3-70B-Magnum-DiamondBF1670.6B131.43 GiB5.31 GiB137.41 GiB6.59 GiB5±8.3%
Trinity-Large-PreviewMoEQ2_K_L399B135.17 GiB1.42 GiB137.16 GiB6.84 GiB29±37%
dots.llm1.instMoEQ6_K143B120.06 GiB16.47 GiB137.10 GiB6.90 GiB13±37%
ERNIE-4.5-300B-A47B-PTQ3_K_M300B132.78 GiB3.59 GiB137.04 GiB6.96 GiB5±8.3%
Trinity-Large-TrueBaseMoEI1-Q2_K399B134.81 GiB1.42 GiB136.81 GiB7.19 GiB29±37%
Hermes-3-Llama-3.1-405BIQ2_M406B127.28 GiB8.37 GiB136.48 GiB7.52 GiB5±8.3%
Llama-3_3-Nemotron-Super-49B-v1_5BF1649.9B92.89 GiB42.50 GiB136.09 GiB7.91 GiB5±8.3%
Valkyrie-49B-v2.1BF1649.9B92.89 GiB42.50 GiB136.09 GiB7.91 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 32,768 context with q8_0 KV cache, the largest being Step-3.7-Flash at UD-Q5_K_M. 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.