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 128K 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 128K context

largest quantization that fits, per model · 2100 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
MiMo-V2.5MoEKV unresolvedUD-IQ4_XS311B139.18 GiB4.22 GiB144.00 GiB0.00 GiB21±37%
Apertus-70B-Instruct-2509BF1670.6B131.51 GiB11.25 GiB143.49 GiB0.51 GiB5±8.3%
command-a-plus-05-2026-bf16MoEQ5_K_S219B141.69 GiB1.24 GiB143.49 GiB0.51 GiB20±37%
Llama-3.3-70B-InstructF1670.6B131.43 GiB11.25 GiB143.35 GiB0.65 GiB5±8.3%
Hermes-4-70BBF1670.6B131.43 GiB11.25 GiB143.35 GiB0.65 GiB5±8.3%
Llama-3.1-70BF1670.6B131.43 GiB11.25 GiB143.35 GiB0.65 GiB5±8.3%
DeepSeek-R1-Distill-Llama-70BF1670.6B131.43 GiB11.25 GiB143.35 GiB0.65 GiB5±8.3%
Athene-70BBF1670.6B131.43 GiB11.25 GiB143.35 GiB0.65 GiB5±8.3%
Hermes-3-Llama-3.1-70BBF1670.6B131.43 GiB11.25 GiB143.35 GiB0.65 GiB5±8.3%
Meta-Llama-3-70B-Instruct-abliterated-v3.5BF1670.6B131.43 GiB11.25 GiB143.35 GiB0.65 GiB5±8.3%
L3.3-70B-Magnum-DiamondBF1670.6B131.43 GiB11.25 GiB143.35 GiB0.65 GiB5±8.3%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ2_K_L402B135.87 GiB6.75 GiB143.19 GiB0.81 GiB22±37%
Step-3.7-FlashUD-Q5_K_S201B128.59 GiB13.79 GiB142.96 GiB1.04 GiB5±8.3%
MiniMax-M2.7MoEQ4_1229B133.65 GiB8.72 GiB142.91 GiB1.09 GiB17±37%
MiniMax-M2.1MoEQ4_1229B133.47 GiB8.72 GiB142.72 GiB1.28 GiB17±37%
MiniMax-M2MoEQ4_1229B133.47 GiB8.72 GiB142.72 GiB1.28 GiB17±37%
MiniMax-M2.5MoEQ4_1229B133.39 GiB8.72 GiB142.64 GiB1.36 GiB17±37%
MiMo-V2-FlashMoEKV unresolvedQ3_K_M310B137.19 GiB4.22 GiB142.00 GiB2.00 GiB21±37%
DeepSeek-V3.1-TerminusMoEIQ1_M685B138.82 GiB2.41 GiB141.86 GiB2.14 GiB23±37%
DeepSeek-V3.2MoEIQ1_M685B138.82 GiB2.41 GiB141.86 GiB2.14 GiB23±37%
cogito-671b-v2.1MoEIQ1_M671B138.82 GiB2.41 GiB141.86 GiB2.14 GiB23±37%
DeepSeek-V3-0324MoEIQ1_M685B138.66 GiB2.41 GiB141.70 GiB2.30 GiB23±37%
r1-1776MoEIQ1_M671B138.66 GiB2.41 GiB141.70 GiB2.30 GiB23±37%
DeepSeek-R1MoEIQ1_M685B138.66 GiB2.41 GiB141.70 GiB2.30 GiB23±37%
GLM-4.7-REAP-218B-A32BMoEQ4_1218B128.16 GiB12.94 GiB141.69 GiB2.31 GiB12±37%
Llama-3_1-Nemotron-51B-InstructQ8_051.5B50.97 GiB90.00 GiB141.66 GiB2.34 GiB5±8.3%
Qwen3.5-REAP-212B-A17BMoEQ5_K_M212B139.97 GiB1.05 GiB141.62 GiB2.38 GiB24±37%
Qwen3.5-397B-A17BMoEUD-IQ3_XXS403B139.55 GiB1.05 GiB141.21 GiB2.79 GiB28±37%
ERNIE-4.5-300B-A47B-PTQ3_K_M300B132.78 GiB7.59 GiB141.05 GiB2.95 GiB5±8.3%
step-3.5-flashQ5_K_S199B126.55 GiB13.79 GiB140.92 GiB3.08 GiB5±8.3%
Qwen3-235B-A22B-Instruct-2507MoEQ4_K_L235B133.28 GiB6.61 GiB140.47 GiB3.53 GiB16±37%
DeepSeek-R1-0528MoEIQ1_M685B137.32 GiB2.41 GiB140.36 GiB3.64 GiB24±37%
DeepSeek-V3.1MoEIQ1_M685B137.32 GiB2.41 GiB140.36 GiB3.64 GiB24±37%
MiniMax-M3MoEIQ2_M427B135.58 GiB4.22 GiB140.36 GiB3.64 GiB21±37%
Qwen3.5-REAP-262B-A17BMoEQ4_K_S262B138.50 GiB1.05 GiB140.16 GiB3.84 GiB26±37%
Llama-3_3-Nemotron-Super-49B-v1_5Q8_049.9B49.36 GiB90.00 GiB140.05 GiB3.95 GiB5±8.3%
Valkyrie-49B-v2.1Q8_049.9B49.36 GiB90.00 GiB140.05 GiB3.95 GiB5±8.3%
Llama-3_3-Nemotron-Super-49B-v1Q8_049.9B49.36 GiB90.00 GiB140.05 GiB3.95 GiB5±8.3%
Qwen3-235B-A22BMoEQ4_K_M235B132.85 GiB6.61 GiB140.04 GiB3.96 GiB16±37%
Qwen3-235B-A22B-Thinking-2507MoEQ4_K_M235B132.85 GiB6.61 GiB140.04 GiB3.96 GiB16±37%
Hy3MoEIQ3_XS299B128.00 GiB11.25 GiB139.83 GiB4.17 GiB15±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ4_K_M236B132.39 GiB6.61 GiB139.58 GiB4.42 GiB16±37%
Qwen3-VL-235B-A22B-InstructMoEQ4_K_M236B132.39 GiB6.61 GiB139.58 GiB4.42 GiB16±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q4_K_M235B132.39 GiB6.61 GiB139.58 GiB4.42 GiB16±37%
Ornith-1.0-397BMoEUD-IQ3_XXS397B137.46 GiB1.05 GiB139.11 GiB4.89 GiB28±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ4_K_M229B128.84 GiB8.72 GiB138.10 GiB5.90 GiB18±37%
Trinity-Large-PreviewMoEQ2_K_L399B135.17 GiB2.33 GiB138.08 GiB5.92 GiB27±37%
Trinity-Large-TrueBaseMoEI1-Q2_K399B134.81 GiB2.33 GiB137.72 GiB6.28 GiB27±37%
dots.llm1.instMoEQ5_K_M143B101.84 GiB34.88 GiB137.30 GiB6.70 GiB8±37%
Yi-1.5-34BF3234.4B128.11 GiB8.44 GiB137.18 GiB6.82 GiB5±8.3%
Devstral-2-123B-Instruct-2512Q8_0125B123.73 GiB12.38 GiB136.81 GiB7.19 GiB5±8.3%
Mistral-Medium-3.5-128BQ8_0128B123.73 GiB12.38 GiB136.81 GiB7.19 GiB5±8.3%
GLM-4.5MoEQ2_K_L358B122.32 GiB12.94 GiB135.85 GiB8.15 GiB14±37%
GLM-4.7MoEQ2_K_L358B122.32 GiB12.94 GiB135.85 GiB8.15 GiB14±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ4_K236B132.67 GiB2.37 GiB135.63 GiB8.37 GiB23±37%
DeepSeek-V2.5MoEQ4_K236B132.67 GiB2.37 GiB135.63 GiB8.37 GiB23±37%
DeepSeek-Coder-V2-InstructMoEQ4_K236B132.67 GiB2.37 GiB135.63 GiB8.37 GiB23±37%
Hermes-4-405BIQ2_S406B117.02 GiB17.72 GiB135.57 GiB8.43 GiB5±8.3%
GLM-4.6MoEQ2_K_L357B121.85 GiB12.94 GiB135.38 GiB8.62 GiB14±37%
Nex-N2-ProMoEIQ2_M397B132.87 GiB1.05 GiB134.53 GiB9.47 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 131,072 context with q4_0 KV cache, the largest being MiMo-V2.5 at UD-IQ4_XS. 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.