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. 2087 of 2118 indexed models fit at 32K 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
vision language 190text 1793audio tts 21image 2audio asr 39video 16embedding 26

What fits at 32K context

largest quantization that fits, per model · 2087 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
gemma-4-26B-A4B-it-Uncensored-MAXMoEF3225.8B94.02 GiB0.82 GiB95.37 GiB0.63 GiB7±8.3%
Step-3.7-FlashQ3_K_M201B87.36 GiB6.92 GiB94.86 GiB1.14 GiB7±8.3%
Mixtral-8x22B-Instruct-v0.1MoEQ5_K_S141B90.32 GiB3.72 GiB94.65 GiB1.35 GiB12±37%
Mixtral-8x22B-v0.1MoEQ5_K_S141B90.31 GiB3.72 GiB94.64 GiB1.36 GiB12±37%
Mixtral-8x22B-v0.1MoEQ5_K_S141B90.31 GiB3.72 GiB94.64 GiB1.36 GiB12±37%
dots.llm1.instMoEQ4_0143B77.44 GiB16.47 GiB94.48 GiB1.52 GiB15±37%
Qwen3.5-122B-A10B-hereticMoEI1-Q6_K123B93.42 GiB0.40 GiB94.40 GiB1.60 GiB36±37%
Llama-3_1-Nemotron-51B-InstructQ8_051.5B50.97 GiB42.50 GiB94.16 GiB1.84 GiB7±8.3%
GLM-4.6-REAP-268B-A32BMoEUD-IQ2_M269B87.26 GiB6.11 GiB93.96 GiB2.04 GiB21±37%
Qwen3-235B-A22B-Instruct-2507MoEIQ3_XS235B90.25 GiB3.12 GiB93.95 GiB2.05 GiB25±37%
Qwen3-235B-A22B-Thinking-2507MoEIQ3_XS235B90.25 GiB3.12 GiB93.95 GiB2.05 GiB25±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q5_K_S139B89.27 GiB4.12 GiB93.92 GiB2.08 GiB25±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-Q5_K_S139B89.27 GiB4.12 GiB93.92 GiB2.08 GiB25±37%
Solar-Open2-250BMoEIQ3_XXS250B89.86 GiB3.19 GiB93.62 GiB2.38 GiB31±37%
Mistral-Small-4-119B-2603MoEUD-Q6_K119B92.60 GiB0.37 GiB93.56 GiB2.44 GiB36±37%
DeepSeek-V4-FlashMoEQ2_K291B92.86 GiB0.03 GiB93.49 GiB2.51 GiB40±37%
MiniMax-M2.7MoEIQ3_XXS229B88.65 GiB4.12 GiB93.30 GiB2.70 GiB28±37%
GLM-4.6VMoEQ6_K_L108B89.58 GiB3.05 GiB93.22 GiB2.78 GiB25±37%
Qwen3-235B-A22B-abliteratedMoEI1-IQ3_XS235B89.36 GiB3.12 GiB93.07 GiB2.93 GiB25±37%
grok-2MoEUD-IQ1_M270B88.02 GiB4.25 GiB92.96 GiB3.04 GiB12±37%
GLM-4.5VMoEI1-Q6_K108B89.28 GiB3.05 GiB92.91 GiB3.09 GiB25±37%
MiMo-V2-FlashMoEKV unresolvedIQ2_M310B90.09 GiB1.99 GiB92.68 GiB3.32 GiB32±37%
Laguna-S-2.1MoEUD-Q6_K118B91.19 GiB0.87 GiB92.63 GiB3.37 GiB33±37%
Qwen3.5-397B-A17BMoEIQ1_M403B91.53 GiB0.50 GiB92.63 GiB3.37 GiB40±37%
Kimi-Linear-48B-A3B-InstructMoEBF1649.1B91.54 GiB0.50 GiB92.60 GiB3.40 GiB7±8.3%
command-a-plus-05-2026-bf16MoEQ3_K_S219B91.28 GiB0.76 GiB92.59 GiB3.41 GiB29±37%
Llama-3_3-Nemotron-Super-49B-v1_5Q8_049.9B49.36 GiB42.50 GiB92.55 GiB3.45 GiB7±8.3%
Valkyrie-49B-v2.1Q8_049.9B49.36 GiB42.50 GiB92.55 GiB3.45 GiB7±8.3%
Llama-3_3-Nemotron-Super-49B-v1Q8_049.9B49.36 GiB42.50 GiB92.55 GiB3.45 GiB7±8.3%
MiMo-V2.5MoEKV unresolvedUD-IQ2_M311B89.93 GiB1.99 GiB92.52 GiB3.48 GiB32±37%
Mistral-Medium-3.5-128BQ5_K_L128B85.78 GiB5.84 GiB92.33 GiB3.67 GiB7±8.3%
Hy3MoEIQ2_S299B86.43 GiB5.31 GiB92.33 GiB3.67 GiB25±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ5_K_L124B90.28 GiB1.46 GiB92.28 GiB3.72 GiB31±37%
GLM-4.7-REAP-218B-A32BMoEIQ3_XS218B85.49 GiB6.11 GiB92.19 GiB3.81 GiB20±37%
MiniMax-M2.1MoEIQ3_XS229B87.32 GiB4.12 GiB91.98 GiB4.02 GiB28±37%
MiniMax-M2MoEIQ3_XS229B87.32 GiB4.12 GiB91.98 GiB4.02 GiB28±37%
Qwen3-VL-235B-A22B-ThinkingMoEUD-IQ3_XXS236B88.22 GiB3.12 GiB91.93 GiB4.07 GiB25±37%
Qwen3-VL-235B-A22B-InstructMoEUD-IQ3_XXS236B88.15 GiB3.12 GiB91.85 GiB4.15 GiB25±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEI1-IQ3_XS229B86.92 GiB4.12 GiB91.57 GiB4.43 GiB28±37%
MiniMax-M2.5MoEI1-IQ3_XS229B86.92 GiB4.12 GiB91.57 GiB4.43 GiB28±37%
step-3.5-flashQ3_K_M199B84.04 GiB6.92 GiB91.54 GiB4.46 GiB7±8.3%
DeepSeek-Coder-V2-Instruct-0724MoEIQ3_XS236B89.69 GiB1.12 GiB91.40 GiB4.60 GiB33±37%
DeepSeek-V2.5MoEIQ3_XS236B89.69 GiB1.12 GiB91.40 GiB4.60 GiB33±37%
DeepSeek-Coder-V2-InstructMoEIQ3_XS236B89.69 GiB1.12 GiB91.40 GiB4.60 GiB33±37%
Dolphin3.0-R1-Mistral-24BF3223.6B87.82 GiB2.66 GiB91.15 GiB4.85 GiB7±8.3%
Dolphin3.0-Mistral-24BF3223.6B87.82 GiB2.66 GiB91.15 GiB4.85 GiB7±8.3%
Mistral-Small-24B-Instruct-2501-abliteratedF3223.6B87.82 GiB2.66 GiB91.15 GiB4.85 GiB7±8.3%
Mistral-Small-24B-Instruct-2501F3223.6B87.82 GiB2.66 GiB91.15 GiB4.85 GiB7±8.3%
Nex-N2-ProMoEIQ1_M397B90.02 GiB0.50 GiB91.12 GiB4.88 GiB40±37%
ERNIE-4.5-300B-A47B-PTUD-IQ1_M300B86.78 GiB3.59 GiB91.04 GiB4.96 GiB7±8.3%
Trinity-Large-PreviewMoEIQ2_XXS399B88.58 GiB1.42 GiB90.57 GiB5.43 GiB38±37%
Trinity-Large-TrueBaseMoEIQ2_XXS399B88.58 GiB1.42 GiB90.57 GiB5.43 GiB38±37%
Hermes-4-405BUD-IQ1_S406B81.23 GiB8.37 GiB90.43 GiB5.57 GiB8±8.3%
GLM-4.5MoEIQ2_XXS358B83.01 GiB6.11 GiB89.71 GiB6.29 GiB23±37%
xLAM-8x7b-rMoEBF1646.7B86.99 GiB2.13 GiB89.71 GiB6.29 GiB13±37%
Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTPQ6_K27.8B88.00 GiB1.06 GiB89.68 GiB6.32 GiB8±8.3%
gpt-oss-20b-hereticMoEQ5_120.9B88.57 GiB0.41 GiB89.51 GiB6.49 GiB22±37%
Huihui-gpt-oss-20b-BF16-abliteratedMoEQ5_120.9B88.57 GiB0.41 GiB89.51 GiB6.49 GiB22±37%
GLM-4.7MoEIQ2_XXS358B82.69 GiB6.11 GiB89.39 GiB6.61 GiB23±37%
GLM-4.6-Derestricted-v3MoEIQ2_XXS357B82.26 GiB6.11 GiB88.96 GiB7.04 GiB23±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?
2087 of 2118 indexed open-weight models fit a Apple M2 Ultra at 32,768 context with q8_0 KV cache, the largest being gemma-4-26B-A4B-it-Uncensored-MAX at F32. 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.