Apple · apple

Apple M2 Ultra

Apple M2 Ultra has 64 GB of unified memory at 819 GB/s — about 44.64 GiB usable after driver and compositor overhead. 2039 of 2118 indexed models fit at 64K context with q4_0 KV. Note only 48 GB of its 64 GB is allocatable to the GPU.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
64 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 1750vision language 185image 2audio tts 21video 16embedding 26audio asr 39

What fits at 64K context

largest quantization that fits, per model · 2039 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Assistant_Pepe_70BQ4_K_L70.6B41.67 GiB5.63 GiB47.97 GiB0.03 GiB14±8.3%
Mistral-Medium-3.5-128BUD-IQ2_M128B41.08 GiB6.19 GiB47.97 GiB0.03 GiB14±8.3%
gemma-2-9b-itF329.2B44.22 GiB3.16 GiB47.97 GiB0.03 GiB14±8.3%
c4ai-command-r-plus-08-2024Q3_K_S104B42.70 GiB4.50 GiB47.93 GiB0.07 GiB14±8.3%
Mistral-MOE-4X7B-Dark-MultiVerse-Uncensored-Enhanced32-24BMoEF1624.2B44.99 GiB2.25 GiB47.83 GiB0.17 GiB8±37%
Qwen3-Coder-NextMoEQ4_K_L79.7B45.60 GiB1.69 GiB47.82 GiB0.18 GiB52±37%
Qwen3-Next-80B-A3B-ThinkingMoEQ4_K_L81.3B45.60 GiB1.69 GiB47.82 GiB0.18 GiB52±37%
Qwen3-Next-80B-A3B-InstructMoEQ4_K_L81.3B45.60 GiB1.69 GiB47.82 GiB0.18 GiB52±37%
HarmonicHarlequin_v5-20BI1-IQ2_M33.3B10.67 GiB36.56 GiB47.82 GiB0.18 GiB14±8.3%
Apertus-70B-Instruct-2509Q4_K_L70.6B41.46 GiB5.63 GiB47.82 GiB0.18 GiB14±8.3%
GLM-4.5VMoEI1-IQ3_XXS108B43.99 GiB3.23 GiB47.80 GiB0.20 GiB37±37%
step-3.5-flashIQ1_M199B40.17 GiB7.04 GiB47.78 GiB0.22 GiB14±8.3%
Qwen3.5-122B-A10B-hereticMoEI1-IQ3_XS123B46.72 GiB0.42 GiB47.72 GiB0.28 GiB57±37%
Step-3.7-FlashIQ1_S201B40.03 GiB7.04 GiB47.65 GiB0.35 GiB14±8.3%
HunyuanImage-2.1Q5_017.5B47.04 GiB0.00 GiB47.64 GiB0.36 GiB14±8.3%
Qwen3.5-88BMoEI1-Q4_K_S87.7B46.63 GiB0.42 GiB47.63 GiB0.37 GiB53±37%
Huihui-Qwen3-Coder-Next-abliteratedMoEQ4_179.7B46.65 GiB0.42 GiB47.61 GiB0.39 GiB61±37%
Meta-Llama-3-70B-InstructQ4_170.6B41.28 GiB5.63 GiB47.58 GiB0.42 GiB14±8.3%
GLM-4.5-Air-REAP-82B-A12BMoEIQ4_XS81.9B43.76 GiB3.23 GiB47.57 GiB0.43 GiB34±37%
Maenad-70BI1-Q4_170.6B41.27 GiB5.63 GiB47.57 GiB0.43 GiB14±8.3%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-ReasonerI1-Q4_170.6B41.27 GiB5.63 GiB47.57 GiB0.43 GiB14±8.3%
L3.3-Electra-R1-70bI1-Q4_170.6B41.27 GiB5.63 GiB47.57 GiB0.43 GiB14±8.3%
L3.3-70B-Magnum-v4-SEQ4_170.6B41.27 GiB5.63 GiB47.57 GiB0.43 GiB14±8.3%
Latxa-Llama-3.1-70B-Instruct-v2I1-Q4_170.6B41.27 GiB5.63 GiB47.57 GiB0.43 GiB14±8.3%
Llama-3.3_70_b_uncensored_continuedI1-Q4_170.6B41.27 GiB5.63 GiB47.57 GiB0.43 GiB14±8.3%
grok-oss-Revenant-70BI1-Q4_170.6B41.27 GiB5.63 GiB47.57 GiB0.43 GiB14±8.3%
Llama-3.1-Nemotron-70B-Instruct-HFI1-Q4_170.6B41.27 GiB5.63 GiB47.57 GiB0.43 GiB14±8.3%
Hermes-4-70B-hereticI1-Q4_170.6B41.27 GiB5.63 GiB47.57 GiB0.43 GiB14±8.3%
Hermes-4-70BQ4_170.6B41.27 GiB5.63 GiB47.57 GiB0.43 GiB14±8.3%
Llama-3.3-70B-Instruct-abliteratedQ4_170.6B41.27 GiB5.63 GiB47.57 GiB0.43 GiB14±8.3%
Llama-3.1-70BQ4_170.6B41.27 GiB5.63 GiB47.57 GiB0.43 GiB14±8.3%
Llama-3.3-70B-InstructQ4_170.6B41.27 GiB5.63 GiB47.57 GiB0.43 GiB14±8.3%
Anubis-70B-v1.2Q4_170.6B41.27 GiB5.63 GiB47.57 GiB0.43 GiB14±8.3%
Golem-70B-v1bI1-Q4_170.6B41.27 GiB5.63 GiB47.57 GiB0.43 GiB14±8.3%
DeepSeek-R1-Distill-Llama-70B-abliteratedI1-Q4_170.6B41.27 GiB5.63 GiB47.57 GiB0.43 GiB14±8.3%
DeepSeek-R1-Distill-Llama-70B-hereticI1-Q4_170.6B41.27 GiB5.63 GiB47.57 GiB0.43 GiB14±8.3%
DeepSeek-R1-Distill-Llama-70BQ4_170.6B41.27 GiB5.63 GiB47.57 GiB0.43 GiB14±8.3%
Legion-V2.1-LLaMa-70BI1-Q4_170.6B41.27 GiB5.63 GiB47.57 GiB0.43 GiB14±8.3%
SEMIKONG-70BQ4_170.6B41.27 GiB5.63 GiB47.57 GiB0.43 GiB14±8.3%
Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16MoEQ5_K_M35.1B46.64 GiB0.35 GiB47.54 GiB0.46 GiB58±37%
Laguna-S-2.1MoEIQ3_XXS118B46.07 GiB0.88 GiB47.53 GiB0.47 GiB52±37%
Llama-4-Scout-17B-16E-Instruct-abliterated-v2MoEKV unresolvedI1-IQ3_S109B43.56 GiB3.38 GiB47.51 GiB0.49 GiB37±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-IQ2_M139B42.58 GiB4.36 GiB47.47 GiB0.53 GiB35±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-IQ2_M139B42.58 GiB4.36 GiB47.47 GiB0.53 GiB35±37%
Devstral-2-123B-Instruct-2512UD-IQ2_M125B40.55 GiB6.19 GiB47.45 GiB0.55 GiB14±8.3%
Mistral-Small-4-119B-2603MoEIQ3_XXS119B46.44 GiB0.40 GiB47.42 GiB0.58 GiB57±37%
Devstral-Small-2-24B-Instruct-2512BF1624.0B43.92 GiB2.81 GiB47.40 GiB0.60 GiB14±8.3%
Magistral-Small-2509BF1624.0B43.92 GiB2.81 GiB47.40 GiB0.60 GiB14±8.3%
Magistral-Small-2507BF1623.6B43.92 GiB2.81 GiB47.40 GiB0.60 GiB14±8.3%
Magistry-24B-v1.1BF1623.6B43.92 GiB2.81 GiB47.40 GiB0.60 GiB14±8.3%
Dolphin3.0-R1-Mistral-24BF1623.6B43.92 GiB2.81 GiB47.40 GiB0.60 GiB14±8.3%
Cydonia_VistralBF1623.6B43.92 GiB2.81 GiB47.40 GiB0.60 GiB14±8.3%
Mistral-Small-3.2-24B-Instruct-2506BF1624.0B43.92 GiB2.81 GiB47.40 GiB0.60 GiB14±8.3%
Mistral-Small-3.2-24B-Instruct-2506-ultra-uncensored-hereticBF1624.0B43.92 GiB2.81 GiB47.40 GiB0.60 GiB14±8.3%
Dans-PersonalityEngine-V1.3.0-24bBF1623.6B43.92 GiB2.81 GiB47.40 GiB0.60 GiB14±8.3%
Devstral-Small-2507BF1623.6B43.92 GiB2.81 GiB47.40 GiB0.60 GiB14±8.3%
Devstral-Small-2505BF1623.6B43.92 GiB2.81 GiB47.40 GiB0.60 GiB14±8.3%
MS3.2-PaintedFantasy-v3-24BBF1623.6B43.92 GiB2.81 GiB47.40 GiB0.60 GiB14±8.3%
Magidonia-24B-v4.3-heretic-v1.2BF1623.6B43.92 GiB2.81 GiB47.40 GiB0.60 GiB14±8.3%
Cogidonia-v2-24BF1623.6B43.92 GiB2.81 GiB47.40 GiB0.60 GiB14±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?
2039 of 2118 indexed open-weight models fit a Apple M2 Ultra at 65,536 context with q4_0 KV cache, the largest being Assistant_Pepe_70B at Q4_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 64 GB, but about 44.64 GiB is available to a model once driver and compositor overhead is accounted for, and only 48 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.