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 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 32K context

largest quantization that fits, per model · 2100 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Trinity-Large-TrueBaseMoEI1-IQ3_XXS399B142.48 GiB0.75 GiB143.81 GiB0.19 GiB29±37%
MiMo-V2.5MoEKV unresolvedUD-IQ4_NL311B142.12 GiB1.05 GiB143.77 GiB0.23 GiB25±37%
Solar-Open2-250BMoEQ4_K_M250B141.39 GiB1.69 GiB143.66 GiB0.34 GiB25±37%
MiniMax-M3MoEQ2_K427B141.46 GiB1.05 GiB143.07 GiB0.93 GiB25±37%
command-a-plus-05-2026-bf16MoEQ5_K_S219B141.69 GiB0.40 GiB142.64 GiB1.36 GiB20±37%
GLM-4.5MoEIQ3_XS358B138.63 GiB3.23 GiB142.45 GiB1.55 GiB20±37%
GLM-4.7MoEIQ3_XS358B138.48 GiB3.23 GiB142.30 GiB1.70 GiB20±37%
Qwen3-Coder-480B-A35B-InstructMoEUD-IQ1_M480B139.43 GiB2.18 GiB142.20 GiB1.80 GiB21±37%
GLM-4.6-Derestricted-v3MoEIQ3_XS357B138.01 GiB3.23 GiB141.83 GiB2.17 GiB20±37%
GLM-4.6MoEIQ3_XS357B138.01 GiB3.23 GiB141.83 GiB2.17 GiB20±37%
Mixtral-8x22B-Instruct-v0.1MoEQ8_0141B139.16 GiB1.97 GiB141.74 GiB2.26 GiB9±37%
Mixtral-8x22B-v0.1MoEQ8_0141B139.16 GiB1.97 GiB141.74 GiB2.26 GiB9±37%
Mixtral-8x22B-v0.1MoEQ8_0141B139.15 GiB1.97 GiB141.73 GiB2.27 GiB9±37%
Qwen3.5-REAP-212B-A17BMoEQ5_K_M212B139.97 GiB0.26 GiB140.83 GiB3.17 GiB25±37%
Step-3.7-FlashUD-Q5_K_M201B136.43 GiB3.67 GiB140.68 GiB3.32 GiB5±8.3%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEQ8_0139B137.78 GiB2.18 GiB140.50 GiB3.50 GiB21±37%
Qwen3.5-397B-A17BMoEUD-IQ3_XXS403B139.55 GiB0.26 GiB140.41 GiB3.59 GiB29±37%
DeepSeek-V3.1-TerminusMoEIQ1_M685B138.82 GiB0.60 GiB140.05 GiB3.95 GiB26±37%
DeepSeek-V3.2MoEIQ1_M685B138.82 GiB0.60 GiB140.05 GiB3.95 GiB26±37%
cogito-671b-v2.1MoEIQ1_M671B138.82 GiB0.60 GiB140.05 GiB3.95 GiB26±37%
DeepSeek-V3-0324MoEIQ1_M685B138.66 GiB0.60 GiB139.89 GiB4.11 GiB26±37%
r1-1776MoEIQ1_M671B138.66 GiB0.60 GiB139.89 GiB4.11 GiB26±37%
DeepSeek-R1MoEIQ1_M685B138.66 GiB0.60 GiB139.89 GiB4.11 GiB26±37%
Qwen3-235B-A22B-Instruct-2507MoEQ4_1235B137.20 GiB1.65 GiB139.43 GiB4.57 GiB20±37%
Qwen3-235B-A22B-Thinking-2507MoEQ4_1235B137.20 GiB1.65 GiB139.43 GiB4.57 GiB20±37%
Qwen3.5-REAP-262B-A17BMoEQ4_K_S262B138.50 GiB0.26 GiB139.36 GiB4.64 GiB27±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ4_1236B137.12 GiB1.65 GiB139.35 GiB4.65 GiB20±37%
Qwen3-VL-235B-A22B-InstructMoEQ4_1236B137.12 GiB1.65 GiB139.35 GiB4.65 GiB20±37%
Qwen3-235B-A22BMoEQ4_1235B137.12 GiB1.65 GiB139.35 GiB4.65 GiB20±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q4_1235B137.12 GiB1.65 GiB139.35 GiB4.65 GiB20±37%
Qwen2.5-72BF1672.7B135.44 GiB2.81 GiB138.93 GiB5.07 GiB5±8.3%
Kimi-Dev-72BBF1672.7B135.44 GiB2.81 GiB138.93 GiB5.07 GiB5±8.3%
Qwen2.5-VL-72B-InstructBF1673.4B135.44 GiB2.81 GiB138.93 GiB5.07 GiB5±8.3%
MiMo-V2-FlashMoEKV unresolvedQ3_K_M310B137.19 GiB1.05 GiB138.84 GiB5.16 GiB26±37%
DeepSeek-R1-0528MoEIQ1_M685B137.32 GiB0.60 GiB138.55 GiB5.45 GiB26±37%
DeepSeek-V3.1MoEIQ1_M685B137.32 GiB0.60 GiB138.55 GiB5.45 GiB26±37%
Ornith-1.0-397BMoEUD-IQ3_XXS397B137.46 GiB0.26 GiB138.32 GiB5.68 GiB30±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ2_K_L402B135.87 GiB1.69 GiB138.13 GiB5.87 GiB30±37%
grok-2MoEIQ4_XS270B134.85 GiB2.25 GiB137.79 GiB6.21 GiB9±37%
GLM-4.6-REAP-268B-A32BMoEIQ4_XS269B133.87 GiB3.23 GiB137.70 GiB6.30 GiB18±37%
Hy3MoEIQ3_M299B133.80 GiB2.81 GiB137.20 GiB6.80 GiB22±37%
Trinity-Large-PreviewMoEQ2_K_L399B135.17 GiB0.75 GiB136.50 GiB7.50 GiB30±37%
MiniMax-M2.7MoEQ4_1229B133.65 GiB2.18 GiB136.37 GiB7.63 GiB24±37%
MiniMax-M2.1MoEQ4_1229B133.47 GiB2.18 GiB136.18 GiB7.82 GiB24±37%
MiniMax-M2MoEQ4_1229B133.47 GiB2.18 GiB136.18 GiB7.82 GiB24±37%
MiniMax-M2.5MoEQ4_1229B133.39 GiB2.18 GiB136.11 GiB7.89 GiB24±37%
ERNIE-4.5-300B-A47B-PTQ3_K_M300B132.78 GiB1.90 GiB135.35 GiB8.65 GiB5±8.3%
step-3.5-flashQ5_K_L199B130.88 GiB3.67 GiB135.13 GiB8.87 GiB5±8.3%
Apertus-70B-Instruct-2509BF1670.6B131.51 GiB2.81 GiB135.06 GiB8.94 GiB5±8.3%
Llama-3.3-70B-InstructF1670.6B131.43 GiB2.81 GiB134.91 GiB9.09 GiB5±8.3%
Hermes-4-70BBF1670.6B131.43 GiB2.81 GiB134.91 GiB9.09 GiB5±8.3%
Llama-3.1-70BF1670.6B131.43 GiB2.81 GiB134.91 GiB9.09 GiB5±8.3%
DeepSeek-R1-Distill-Llama-70BF1670.6B131.43 GiB2.81 GiB134.91 GiB9.09 GiB5±8.3%
Athene-70BBF1670.6B131.43 GiB2.81 GiB134.91 GiB9.09 GiB5±8.3%
Hermes-3-Llama-3.1-70BBF1670.6B131.43 GiB2.81 GiB134.91 GiB9.09 GiB5±8.3%
Meta-Llama-3-70B-Instruct-abliterated-v3.5BF1670.6B131.43 GiB2.81 GiB134.91 GiB9.09 GiB5±8.3%
L3.3-70B-Magnum-DiamondBF1670.6B131.43 GiB2.81 GiB134.91 GiB9.09 GiB5±8.3%
DeepSeek-Coder-V2-Instruct-0724MoEQ4_K236B132.67 GiB0.59 GiB133.85 GiB10.15 GiB26±37%
DeepSeek-V2.5MoEQ4_K236B132.67 GiB0.59 GiB133.85 GiB10.15 GiB26±37%
DeepSeek-Coder-V2-InstructMoEQ4_K236B132.67 GiB0.59 GiB133.85 GiB10.15 GiB26±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 32,768 context with q4_0 KV cache, the largest being Trinity-Large-TrueBase at I1-IQ3_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.