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. 2079 of 2118 indexed models fit at 128K context with f16 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 1784vision language 191image 2audio tts 21audio asr 39embedding 26video 16

What fits at 128K context

largest quantization that fits, per model · 2079 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
GLM-4.7-REAP-218B-A32BMoEQ3_K_L218B97.31 GiB46.00 GiB143.90 GiB0.10 GiB7±37%
Trinity-Large-PreviewMoEQ2_K399B135.04 GiB8.29 GiB143.90 GiB0.10 GiB19±37%
Qwen3.5-397B-A17BMoEUD-IQ3_XXS403B139.55 GiB3.75 GiB143.90 GiB0.10 GiB24±37%
Trinity-Large-TrueBaseMoEI1-Q2_K399B134.81 GiB8.29 GiB143.68 GiB0.32 GiB19±37%
GPT-NeoX-20B-ErebusI1-Q3_K_L20.6B10.96 GiB132.00 GiB143.60 GiB0.40 GiB5±8.3%
Delphi-25B-SimpleRL-MathI1-IQ3_XXS25.0B9.03 GiB133.88 GiB143.53 GiB0.47 GiB5±8.3%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedUD-IQ1_M402B118.78 GiB24.00 GiB143.36 GiB0.64 GiB12±37%
Qwen3.5-REAP-262B-A17BMoEQ4_K_S262B138.50 GiB3.75 GiB142.85 GiB1.15 GiB22±37%
grok-2MoEQ3_K_S270B109.94 GiB32.00 GiB142.63 GiB1.37 GiB6±37%
GLM-4.5MoEIQ2_S358B95.61 GiB46.00 GiB142.20 GiB1.80 GiB7±37%
Step-3.5-Flash-REAP-121B-A11BI1-Q6_K121B92.51 GiB49.03 GiB142.12 GiB1.88 GiB5±8.3%
Qwen3-235B-A22B-abliteratedMoEIQ4_XS235B117.97 GiB23.50 GiB142.05 GiB1.95 GiB10±37%
GLM-4.7MoEIQ2_S358B95.29 GiB46.00 GiB141.88 GiB2.12 GiB7±37%
Hy3MoEQ2_K299B101.28 GiB40.00 GiB141.87 GiB2.13 GiB8±37%
Ornith-1.0-397BMoEUD-IQ3_XXS397B137.46 GiB3.75 GiB141.81 GiB2.19 GiB24±37%
MiniMax-M2.1MoEI1-Q3_K_L229B110.22 GiB31.00 GiB141.76 GiB2.24 GiB9±37%
MiniMax-M2.5MoEI1-Q3_K_L229B110.22 GiB31.00 GiB141.76 GiB2.24 GiB9±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ3_K_L229B110.22 GiB31.00 GiB141.76 GiB2.24 GiB9±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ4_K236B132.67 GiB8.44 GiB141.70 GiB2.30 GiB17±37%
DeepSeek-V2.5MoEQ4_K236B132.67 GiB8.44 GiB141.70 GiB2.30 GiB17±37%
DeepSeek-Coder-V2-InstructMoEQ4_K236B132.67 GiB8.44 GiB141.70 GiB2.30 GiB17±37%
GLM-4.6-Derestricted-v3MoEIQ2_S357B94.86 GiB46.00 GiB141.45 GiB2.55 GiB7±37%
GLM-4.6MoEIQ2_S357B94.86 GiB46.00 GiB141.45 GiB2.55 GiB7±37%
MiMo-V2.5MoEKV unresolvedQ3_K_S311B125.83 GiB15.00 GiB141.43 GiB2.57 GiB14±37%
Qwen3-235B-A22B-Instruct-2507MoEIQ4_XS235B117.24 GiB23.50 GiB141.32 GiB2.68 GiB10±37%
Qwen3-235B-A22B-Thinking-2507MoEIQ4_XS235B117.24 GiB23.50 GiB141.32 GiB2.68 GiB10±37%
MiMo-V2-FlashMoEKV unresolvedI1-IQ3_M310B125.52 GiB15.00 GiB141.12 GiB2.88 GiB14±37%
Qwen3-235B-A22BMoEIQ4_XS235B116.89 GiB23.50 GiB140.97 GiB3.03 GiB10±37%
Qwen3-VL-235B-A22B-ThinkingMoEIQ4_XS236B116.70 GiB23.50 GiB140.78 GiB3.22 GiB10±37%
Qwen3-VL-235B-A22B-InstructMoEIQ4_XS236B116.70 GiB23.50 GiB140.78 GiB3.22 GiB10±37%
MiniMax-M3MoEUD-IQ2_M427B124.99 GiB15.00 GiB140.56 GiB3.44 GiB14±37%
llama2-22b-chat-wizard-uncensoredQ3_K_M21.8B9.88 GiB130.00 GiB140.50 GiB3.50 GiB5±8.3%
Step-3.7-FlashUD-IQ4_NL201B90.63 GiB49.03 GiB140.24 GiB3.76 GiB5±8.3%
Devstral-2-123B-Instruct-2512Q6_K125B95.53 GiB44.00 GiB140.24 GiB3.76 GiB5±8.3%
DeepSeek-R1MoEUD-IQ1_S685B130.60 GiB8.58 GiB139.81 GiB4.19 GiB18±37%
Qwen3-Coder-REAP-363B-A35BMoEUD-IQ1_M363B107.85 GiB31.00 GiB139.43 GiB4.57 GiB9±37%
Trinity-Large-ThinkingMoEQ2_K_L399B130.22 GiB8.29 GiB139.09 GiB4.91 GiB20±37%
GLM-4.6-REAP-268B-A32BMoEQ2_K_L269B92.11 GiB46.00 GiB138.70 GiB5.30 GiB7±37%
Behemoth-X-123B-v2Q6_K123B93.68 GiB44.00 GiB138.38 GiB5.62 GiB5±8.3%
Mistral-Large-Instruct-2411Q6_K123B93.68 GiB44.00 GiB138.38 GiB5.62 GiB5±8.3%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q6_K139B106.40 GiB31.00 GiB137.94 GiB6.06 GiB9±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEQ6_K139B106.40 GiB31.00 GiB137.94 GiB6.06 GiB9±37%
Nex-N2-ProMoEIQ2_M397B132.87 GiB3.75 GiB137.22 GiB6.78 GiB24±37%
v6-Finch-14B-HFQ8_014.1B14.55 GiB122.00 GiB137.14 GiB6.86 GiB5±8.3%
Solar-Open2-250BMoEQ3_K_M250B111.63 GiB24.00 GiB136.21 GiB7.79 GiB11±37%
Mixtral-8x22B-v0.1MoEQ6_K141B107.60 GiB28.00 GiB136.21 GiB7.79 GiB6±37%
SuperHY3-abliterated-NVFP4MoEIQ2_M172B95.10 GiB40.00 GiB135.69 GiB8.31 GiB8±37%
c4ai-command-r-plus-08-2024Q8_0104B102.74 GiB32.00 GiB135.47 GiB8.53 GiB5±8.3%
MiniMax-M2.7MoEUD-IQ4_NL229B103.15 GiB31.00 GiB134.68 GiB9.32 GiB9±37%
command-a-plus-05-2026-bf16MoEQ4_1219B129.29 GiB4.42 GiB134.27 GiB9.73 GiB18±37%
step-3.5-flashQ3_K_M199B84.04 GiB49.03 GiB133.64 GiB10.36 GiB5±8.3%
DeepSeek-V3-0324MoEIQ1_S685B124.38 GiB8.58 GiB133.59 GiB10.41 GiB18±37%
MiniMax-M2MoEQ3_K_M229B101.77 GiB31.00 GiB133.30 GiB10.70 GiB9±37%
GLM-4.5-AirMoEQ8_0110B109.39 GiB23.00 GiB132.97 GiB11.03 GiB10±37%
GLM-4.5-Air-DerestrictedMoEQ8_0110B109.39 GiB23.00 GiB132.97 GiB11.03 GiB10±37%
Qwen3.6-35B-A3B-abliterated-MAXMoEF3235.1B129.13 GiB2.50 GiB132.19 GiB11.81 GiB24±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ8_0109B106.67 GiB24.00 GiB131.24 GiB12.76 GiB10±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ8_0124B119.65 GiB11.00 GiB131.20 GiB12.80 GiB16±37%
Mistral-Medium-3.5-128BQ5_K_L128B85.78 GiB44.00 GiB130.49 GiB13.51 GiB5±8.3%
ERNIE-4.5-300B-A47B-PTQ2_K_L300B101.76 GiB27.00 GiB129.44 GiB14.56 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?
2079 of 2118 indexed open-weight models fit a Apple M2 Ultra at 131,072 context with f16 KV cache, the largest being GLM-4.7-REAP-218B-A32B at Q3_K_L. 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.