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. 2077 of 2118 indexed models fit at 64K context with f16 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 1783vision language 190audio tts 21image 2audio asr 39video 16embedding 26

What fits at 64K context

largest quantization that fits, per model · 2077 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Devstral-2-123B-Instruct-2512Q4_1125B73.08 GiB22.00 GiB95.79 GiB0.21 GiB7±8.3%
Mistral-Medium-3.5-128BI1-Q4_1128B73.08 GiB22.00 GiB95.79 GiB0.21 GiB7±8.3%
XORTRON-NXTXPRTXXLI1-Q4_1128B73.08 GiB22.00 GiB95.79 GiB0.21 GiB7±8.3%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ6_K_L109B83.13 GiB12.00 GiB95.71 GiB0.29 GiB17±37%
xLAM-8x7b-rMoEBF1646.7B86.99 GiB8.00 GiB95.58 GiB0.42 GiB11±37%
Qwen3.5-122B-A10B-hereticMoEI1-Q6_K123B93.42 GiB1.50 GiB95.50 GiB0.50 GiB33±37%
MiMo-V2-FlashMoEKV unresolvedUD-IQ1_M310B87.36 GiB7.50 GiB95.46 GiB0.54 GiB23±37%
command-r-35b-writer-v2I1-IQ3_S35.0B14.77 GiB80.00 GiB95.43 GiB0.57 GiB7±8.3%
GLM-4.5MoEIQ1_S358B71.81 GiB23.00 GiB95.40 GiB0.60 GiB12±37%
GLM-4.7MoEIQ1_S358B71.50 GiB23.00 GiB95.09 GiB0.91 GiB12±37%
Laguna-S-2.1MoEUD-Q6_K118B91.19 GiB3.14 GiB94.90 GiB1.10 GiB28±37%
MiMo-V2.5MoEKV unresolvedIQ2_XS311B86.68 GiB7.50 GiB94.78 GiB1.22 GiB23±37%
ERNIE-4.5-300B-A47B-PTUD-IQ1_S300B80.54 GiB13.50 GiB94.72 GiB1.28 GiB7±8.3%
GLM-4.6-Derestricted-v3MoEIQ1_S357B71.07 GiB23.00 GiB94.66 GiB1.34 GiB12±37%
GLM-4.6MoEIQ1_S357B71.07 GiB23.00 GiB94.66 GiB1.34 GiB12±37%
Mistral-Small-4-119B-2603MoEUD-Q6_K119B92.60 GiB1.41 GiB94.59 GiB1.41 GiB33±37%
DeepSeek-Coder-V2-Instruct-0724MoEIQ3_XS236B89.69 GiB4.22 GiB94.50 GiB1.50 GiB26±37%
DeepSeek-V2.5MoEIQ3_XS236B89.69 GiB4.22 GiB94.50 GiB1.50 GiB26±37%
DeepSeek-Coder-V2-InstructMoEIQ3_XS236B89.69 GiB4.22 GiB94.50 GiB1.50 GiB26±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q4_K_M139B78.40 GiB15.50 GiB94.44 GiB1.56 GiB15±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-Q4_K_M139B78.40 GiB15.50 GiB94.44 GiB1.56 GiB15±37%
GLM-4.7-REAP-218B-A32BMoEUD-IQ2_M218B70.78 GiB23.00 GiB94.37 GiB1.63 GiB11±37%
Mixtral-8x22B-Instruct-v0.1MoEQ4_K_M141B79.71 GiB14.00 GiB94.33 GiB1.67 GiB10±37%
Mixtral-8x22B-v0.1MoEQ4_K_M141B79.71 GiB14.00 GiB94.32 GiB1.68 GiB10±37%
Mixtral-8x22B-v0.1MoEQ4_K_M141B79.71 GiB14.00 GiB94.32 GiB1.68 GiB10±37%
command-a-plus-05-2026-bf16MoEQ3_K_S219B91.28 GiB2.42 GiB94.26 GiB1.74 GiB26±37%
Behemoth-X-123B-v2Q4_1123B71.45 GiB22.00 GiB94.15 GiB1.85 GiB7±8.3%
Step-3.7-FlashUD-IQ3_XXS201B68.54 GiB25.03 GiB94.15 GiB1.85 GiB7±8.3%
Qwen3.5-397B-A17BMoEIQ1_M403B91.53 GiB1.88 GiB94.01 GiB1.99 GiB35±37%
Kimi-Linear-48B-A3B-InstructMoEBF1649.1B91.54 GiB1.90 GiB93.99 GiB2.01 GiB7±8.3%
MiniMax-M2.7MoEUD-IQ3_S229B77.87 GiB15.50 GiB93.91 GiB2.09 GiB16±37%
MiniMax-M2.1MoEQ2_K_L229B77.72 GiB15.50 GiB93.75 GiB2.25 GiB16±37%
MiniMax-M2.5MoEQ2_K_L229B77.72 GiB15.50 GiB93.75 GiB2.25 GiB16±37%
MiniMax-M2MoEQ2_K_L229B77.72 GiB15.50 GiB93.75 GiB2.25 GiB16±37%
Trinity-Large-PreviewMoEIQ2_XXS399B88.58 GiB4.54 GiB93.69 GiB2.31 GiB30±37%
Trinity-Large-TrueBaseMoEIQ2_XXS399B88.58 GiB4.54 GiB93.69 GiB2.31 GiB30±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEI1-Q2_K229B77.58 GiB15.50 GiB93.62 GiB2.38 GiB16±37%
DeepSeek-V4-FlashMoEQ2_K291B92.86 GiB0.06 GiB93.52 GiB2.48 GiB40±37%
Step-3.5-Flash-REAP-121B-A11BI1-Q4_K_M121B67.82 GiB25.03 GiB93.43 GiB2.57 GiB7±8.3%
Noromaid-20b-v0.1.1I1-Q6_K20.0B15.28 GiB77.50 GiB93.37 GiB2.63 GiB7±8.3%
Nethena-20BQ6_K20.0B15.28 GiB77.50 GiB93.37 GiB2.63 GiB7±8.3%
Qwen2.5-72B-InstructQ8_072.7B72.21 GiB20.00 GiB92.89 GiB3.11 GiB7±8.3%
grok-2MoEUD-TQ1_0270B76.17 GiB16.00 GiB92.86 GiB3.14 GiB10±37%
GLM-4.6-REAP-268B-A32BMoEUD-IQ1_S269B69.09 GiB23.00 GiB92.68 GiB3.32 GiB12±37%
HuatuoGPT-o1-72BQ8_072.7B71.96 GiB20.00 GiB92.64 GiB3.36 GiB7±8.3%
Qwen2.5-72B-Instruct-abliteratedQ8_072.7B71.96 GiB20.00 GiB92.64 GiB3.36 GiB7±8.3%
MiroThinker-v1.0-72BQ8_072.7B71.96 GiB20.00 GiB92.64 GiB3.36 GiB7±8.3%
Qwen2.5-Math-72B-InstructQ8_072.7B71.96 GiB20.00 GiB92.64 GiB3.36 GiB7±8.3%
Qwen2.5-72BQ8_072.7B71.96 GiB20.00 GiB92.64 GiB3.36 GiB7±8.3%
Rombo-LLM-V3.0-Qwen-72bQ8_072.7B71.96 GiB20.00 GiB92.64 GiB3.36 GiB7±8.3%
EVA-Qwen2.5-72B-v0.2Q8_072.7B71.96 GiB20.00 GiB92.64 GiB3.36 GiB7±8.3%
Kimi-Dev-72BQ8_072.7B71.96 GiB20.00 GiB92.64 GiB3.36 GiB7±8.3%
Chuluun-Qwen2.5-72B-v0.01Q8_072.7B71.96 GiB20.00 GiB92.64 GiB3.36 GiB7±8.3%
magnum-v4-72bQ8_072.7B71.96 GiB20.00 GiB92.64 GiB3.36 GiB7±8.3%
KAT-Dev-72B-ExpQ8_072.7B71.96 GiB20.00 GiB92.64 GiB3.36 GiB7±8.3%
Homer-v1.0-Qwen2.5-72BQ8_072.7B71.96 GiB20.00 GiB92.64 GiB3.36 GiB7±8.3%
Qwen2.5-VL-72B-InstructQ8_073.4B71.96 GiB20.00 GiB92.64 GiB3.36 GiB7±8.3%
Tower-Plus-72B-ultra-uncensored-hereticQ8_072.7B71.96 GiB20.00 GiB92.64 GiB3.36 GiB7±8.3%
Chronos-Platinum-72BQ8_072.7B71.96 GiB20.00 GiB92.64 GiB3.36 GiB7±8.3%
UI-TARS-72B-DPOQ8_073.4B71.96 GiB20.00 GiB92.64 GiB3.36 GiB7±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?
2077 of 2118 indexed open-weight models fit a Apple M2 Ultra at 65,536 context with f16 KV cache, the largest being Devstral-2-123B-Instruct-2512 at Q4_1. 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.