Apple · apple

Apple M2 Ultra

Apple M2 Ultra has 128 GB of unified memory at 819 GB/s — about 89.28 GiB usable after driver and compositor overhead. 2090 of 2118 indexed models fit at 8K context with q8_0 KV. Note only 96 GB of its 128 GB is allocatable to the GPU.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
128 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 1795vision language 191audio tts 21image 2audio asr 39video 16embedding 26

What fits at 8K context

largest quantization that fits, per model · 2090 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Qwen3-Coder-REAP-363B-A35BMoEUD-TQ1_0363B94.37 GiB1.03 GiB95.98 GiB0.02 GiB27±37%
ERNIE-4.5-300B-A47B-PTUD-IQ2_XXS300B94.32 GiB0.90 GiB95.89 GiB0.11 GiB7±8.3%
Qwen3-235B-A22B-abliteratedMoEI1-IQ3_S235B94.50 GiB0.78 GiB95.87 GiB0.13 GiB28±37%
MiniMax-M2.7MoEUD-Q3_K_M229B94.29 GiB1.03 GiB95.86 GiB0.14 GiB34±37%
Behemoth-X-123B-v2Q6_K123B93.68 GiB1.46 GiB95.84 GiB0.16 GiB7±8.3%
Mistral-Large-Instruct-2411Q6_K123B93.68 GiB1.46 GiB95.84 GiB0.16 GiB7±8.3%
Qwen3-VL-235B-A22B-ThinkingMoEQ3_K_S236B94.48 GiB0.78 GiB95.84 GiB0.16 GiB28±37%
Qwen3-VL-235B-A22B-InstructMoEQ3_K_S236B94.48 GiB0.78 GiB95.84 GiB0.16 GiB28±37%
Qwen3-235B-A22BMoEQ3_K_S235B94.48 GiB0.78 GiB95.84 GiB0.16 GiB28±37%
Laguna-S-2.1MoEQ6_K_L118B94.83 GiB0.27 GiB95.68 GiB0.32 GiB34±37%
MiMo-V2-FlashMoEKV unresolvedUD-IQ2_XXS310B94.58 GiB0.50 GiB95.68 GiB0.32 GiB35±37%
DeepSeek-Coder-V2-Instruct-0724MoEIQ3_S236B94.70 GiB0.28 GiB95.57 GiB0.43 GiB34±37%
DeepSeek-V2.5MoEIQ3_S236B94.70 GiB0.28 GiB95.57 GiB0.43 GiB34±37%
DeepSeek-Coder-V2-InstructMoEIQ3_S236B94.70 GiB0.28 GiB95.57 GiB0.43 GiB34±37%
Step-3.5-Flash-REAP-121B-A11BI1-Q6_K121B92.51 GiB2.14 GiB95.23 GiB0.77 GiB7±8.3%
grok-2MoEQ2_K_L270B93.41 GiB1.06 GiB95.17 GiB0.83 GiB13±37%
MiniMax-M3MoEIQ1_M427B93.82 GiB0.50 GiB94.88 GiB1.12 GiB35±37%
gemma-4-26B-A4B-it-Uncensored-MAXMoEF3225.8B94.02 GiB0.32 GiB94.87 GiB1.13 GiB7±8.3%
MiniMax-M2.1MoEI1-IQ3_M229B93.13 GiB1.03 GiB94.69 GiB1.31 GiB34±37%
MiniMax-M2.5MoEI1-IQ3_M229B93.13 GiB1.03 GiB94.69 GiB1.31 GiB34±37%
Mixtral-8x22B-Instruct-v0.1MoEQ5_K_M141B93.11 GiB0.93 GiB94.65 GiB1.35 GiB13±37%
Mixtral-8x22B-v0.1MoEQ5_K_M141B93.11 GiB0.93 GiB94.65 GiB1.35 GiB13±37%
dots.llm1.instMoEQ4_K_L143B89.95 GiB4.12 GiB94.64 GiB1.36 GiB25±37%
Mixtral-8x22B-v0.1MoEQ5_K_M141B93.10 GiB0.93 GiB94.64 GiB1.36 GiB13±37%
GLM-4.6-REAP-268B-A32BMoEQ2_K_L269B92.11 GiB1.53 GiB94.23 GiB1.77 GiB26±37%
Qwen3.5-122B-A10B-hereticMoEI1-Q6_K123B93.42 GiB0.10 GiB94.10 GiB1.90 GiB37±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedTQ1_0402B92.59 GiB0.80 GiB93.97 GiB2.03 GiB42±37%
MiniMax-M2MoEQ3_K_S229B92.31 GiB1.03 GiB93.87 GiB2.13 GiB34±37%
GLM-4.5-Air-DerestrictedMoEQ6_K110B92.37 GiB0.76 GiB93.71 GiB2.29 GiB28±37%
GLM-4.5-AirMoEQ6_K110B92.37 GiB0.76 GiB93.71 GiB2.29 GiB28±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q5_K_M139B91.98 GiB1.03 GiB93.54 GiB2.46 GiB30±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-Q5_K_M139B91.98 GiB1.03 GiB93.54 GiB2.46 GiB30±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEI1-IQ3_S229B91.94 GiB1.03 GiB93.50 GiB2.50 GiB34±37%
DeepSeek-V4-FlashMoEQ2_K291B92.86 GiB0.03 GiB93.49 GiB2.51 GiB40±37%
Step-3.7-FlashUD-IQ4_NL201B90.63 GiB2.14 GiB93.35 GiB2.65 GiB7±8.3%
Mistral-Small-4-119B-2603MoEUD-Q6_K119B92.60 GiB0.09 GiB93.28 GiB2.72 GiB37±37%
GLM-4.7MoEUD-IQ1_S358B90.50 GiB1.53 GiB92.62 GiB3.38 GiB29±37%
GLM-4.7-REAP-218B-A32BMoEQ3_K_S218B90.39 GiB1.53 GiB92.50 GiB3.50 GiB24±37%
GLM-4.5MoEUD-IQ1_S358B90.38 GiB1.53 GiB92.50 GiB3.50 GiB29±37%
GLM-4.6MoEUD-IQ1_S357B90.28 GiB1.53 GiB92.40 GiB3.60 GiB30±37%
Qwen3.5-397B-A17BMoEIQ1_M403B91.53 GiB0.12 GiB92.26 GiB3.74 GiB41±37%
Kimi-Linear-48B-A3B-InstructMoEBF1649.1B91.54 GiB0.13 GiB92.22 GiB3.78 GiB7±8.3%
command-a-plus-05-2026-bf16MoEQ3_K_S219B91.28 GiB0.36 GiB92.20 GiB3.80 GiB29±37%
Qwen3-235B-A22B-Instruct-2507MoEIQ3_XS235B90.25 GiB0.78 GiB91.61 GiB4.39 GiB29±37%
Qwen3-235B-A22B-Thinking-2507MoEIQ3_XS235B90.25 GiB0.78 GiB91.61 GiB4.39 GiB29±37%
Solar-Open2-250BMoEIQ3_XXS250B89.86 GiB0.80 GiB91.23 GiB4.77 GiB37±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ5_K_L124B90.28 GiB0.37 GiB91.19 GiB4.81 GiB34±37%
Hermes-4-405BUD-IQ1_M406B88.23 GiB2.09 GiB91.15 GiB4.85 GiB7±8.3%
MiMo-V2.5MoEKV unresolvedUD-IQ2_M311B89.93 GiB0.50 GiB91.02 GiB4.98 GiB37±37%
GLM-4.6VMoEQ6_K_L108B89.58 GiB0.76 GiB90.93 GiB5.07 GiB29±37%
Nex-N2-ProMoEIQ1_M397B90.02 GiB0.12 GiB90.74 GiB5.26 GiB42±37%
GLM-4.5VMoEI1-Q6_K108B89.28 GiB0.76 GiB90.62 GiB5.38 GiB29±37%
Hermes-3-Llama-3.1-405BIQ1_M406B87.08 GiB2.09 GiB90.00 GiB6.00 GiB8±8.3%
Trinity-Large-PreviewMoEIQ2_XXS399B88.58 GiB0.67 GiB89.82 GiB6.18 GiB41±37%
Trinity-Large-TrueBaseMoEIQ2_XXS399B88.58 GiB0.67 GiB89.82 GiB6.18 GiB41±37%
gpt-oss-20b-hereticMoEQ5_120.9B88.57 GiB0.11 GiB89.21 GiB6.79 GiB22±37%
Huihui-gpt-oss-20b-BF16-abliteratedMoEQ5_120.9B88.57 GiB0.11 GiB89.21 GiB6.79 GiB22±37%
Dolphin3.0-R1-Mistral-24BF3223.6B87.82 GiB0.66 GiB89.15 GiB6.85 GiB8±8.3%
Dolphin3.0-Mistral-24BF3223.6B87.82 GiB0.66 GiB89.15 GiB6.85 GiB8±8.3%
Mistral-Small-24B-Instruct-2501-abliteratedF3223.6B87.82 GiB0.66 GiB89.15 GiB6.85 GiB8±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?
2090 of 2118 indexed open-weight models fit a Apple M2 Ultra at 8,192 context with q8_0 KV cache, the largest being Qwen3-Coder-REAP-363B-A35B at UD-TQ1_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 128 GB, but about 89.28 GiB is available to a model once driver and compositor overhead is accounted for, and only 96 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.