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. 1988 of 2118 indexed models fit at 128K context with q8_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 1701vision language 183image 2audio tts 21audio asr 39video 16embedding 26

What fits at 128K context

largest quantization that fits, per model · 1988 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Apertus-70B-Instruct-2509UD-IQ3_XXS70.6B26.01 GiB21.25 GiB47.99 GiB0.01 GiB14±8.3%
glm-4-9b-chat-abliteratedIQ4_XS9.4B4.89 GiB42.50 GiB47.99 GiB0.01 GiB14±8.3%
Rombo-LLM-V3.0-Qwen-72bI1-IQ2_S72.7B26.02 GiB21.25 GiB47.95 GiB0.05 GiB14±8.3%
Qwen2.5-72B-Instruct-abliteratedI1-IQ2_S72.7B26.02 GiB21.25 GiB47.95 GiB0.05 GiB14±8.3%
Qwen2.5-72B-Instruct-abliterated-v2I1-IQ2_S72.7B26.02 GiB21.25 GiB47.95 GiB0.05 GiB14±8.3%
HuatuoGPT-o1-72BIQ2_S72.7B26.02 GiB21.25 GiB47.95 GiB0.05 GiB14±8.3%
MiroThinker-v1.0-72BI1-IQ2_S72.7B26.02 GiB21.25 GiB47.95 GiB0.05 GiB14±8.3%
Malaysian-Qwen2.5-72B-InstructI1-IQ2_S72.7B26.02 GiB21.25 GiB47.95 GiB0.05 GiB14±8.3%
Qwen2.5-72BI1-IQ2_S72.7B26.02 GiB21.25 GiB47.95 GiB0.05 GiB14±8.3%
magnum-v4-72bI1-IQ2_S72.7B26.02 GiB21.25 GiB47.95 GiB0.05 GiB14±8.3%
KAT-Dev-72B-ExpIQ2_S72.7B26.02 GiB21.25 GiB47.95 GiB0.05 GiB14±8.3%
Homer-v1.0-Qwen2.5-72BIQ2_S72.7B26.02 GiB21.25 GiB47.95 GiB0.05 GiB14±8.3%
Tower-Plus-72B-ultra-uncensored-hereticI1-IQ2_S72.7B26.02 GiB21.25 GiB47.95 GiB0.05 GiB14±8.3%
Qwen2.5-VL-72B-InstructIQ2_S73.4B26.02 GiB21.25 GiB47.95 GiB0.05 GiB14±8.3%
UI-TARS-72B-DPOIQ2_S73.4B26.02 GiB21.25 GiB47.95 GiB0.05 GiB14±8.3%
GLM-4.6VMoEIQ2_XXS108B35.11 GiB12.22 GiB47.90 GiB0.10 GiB21±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedIQ2_M109B34.56 GiB12.75 GiB47.89 GiB0.11 GiB21±37%
Hunyuan-A13B-InstructMoEQ3_K_L80.4B38.84 GiB8.50 GiB47.89 GiB0.11 GiB14±8.3%
glm-4-9b-chat-1mQ3_K_M9.5B4.76 GiB42.50 GiB47.86 GiB0.14 GiB14±8.3%
codegeex4-all-9bQ3_K_M9.4B4.72 GiB42.50 GiB47.81 GiB0.19 GiB14±8.3%
glm-4-9b-chatQ3_K_M9.4B4.72 GiB42.50 GiB47.81 GiB0.19 GiB14±8.3%
Mistral-Small-Instruct-2409Q2_K22.2B32.27 GiB14.88 GiB47.75 GiB0.25 GiB14±8.3%
L3-DARKEST-PLANET-16.5BIQ4_XS16.5B28.30 GiB18.86 GiB47.75 GiB0.25 GiB14±8.3%
Llama-4-Scout-17B-16E-Instruct-abliterated-v2MoEKV unresolvedI1-Q2_K_S109B34.42 GiB12.75 GiB47.75 GiB0.25 GiB21±37%
Assistant_Pepe_70BQ2_K70.6B25.79 GiB21.25 GiB47.72 GiB0.28 GiB14±8.3%
Hermes-4-70BUD-IQ3_XXS70.6B25.76 GiB21.25 GiB47.68 GiB0.32 GiB14±8.3%
Llama-3.3-70B-InstructUD-IQ3_XXS70.6B25.76 GiB21.25 GiB47.68 GiB0.32 GiB14±8.3%
DeepSeek-R1-Distill-Llama-70BUD-IQ3_XXS70.6B25.76 GiB21.25 GiB47.68 GiB0.32 GiB14±8.3%
HunyuanImage-2.1Q5_017.5B47.04 GiB0.00 GiB47.64 GiB0.36 GiB14±8.3%
Huihui-Qwen3-Coder-Next-abliteratedMoEQ4_K_L79.7B45.44 GiB1.59 GiB47.57 GiB0.43 GiB53±37%
Ace-Step1.5Q8_0160M43.77 GiB3.21 GiB47.52 GiB0.48 GiB14±8.3%
Maenad-70BI1-IQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-ReasonerI1-IQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
Rombos-LLM-70b-Llama-3.3I1-IQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
L3.3-Electra-R1-70bI1-IQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
L3.3-70B-Magnum-v4-SEIQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
Latxa-Llama-3.1-70B-Instruct-v2I1-IQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
Llama-3.3_70_b_uncensored_continuedI1-IQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
Llama-3.3-70B-Instruct-abliteratedI1-IQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
grok-oss-Revenant-70BI1-IQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
Llama-3.1-Nemotron-70B-Instruct-HFI1-IQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
L3.3-70B-Euryale-v2.3I1-IQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
Hermes-3-Llama-3.1-70BIQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
Hermes-4-70B-hereticI1-IQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
Llama-3.1-70BIQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
Anubis-70B-v1.2IQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
Golem-70B-v1bI1-IQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
DeepSeek-R1-Distill-Llama-70B-abliteratedI1-IQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
DeepSeek-R1-Distill-Llama-70B-hereticI1-IQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
Legion-V2.1-LLaMa-70BI1-IQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
Tess-R1-Limerick-Llama-3.1-70BIQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
SEMIKONG-70BIQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
functionary-medium-v3.2KV unresolvedIQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
Llama-3.1-WhiteRabbitNeo-2-70BIQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
Infinity-Instruct-7M-Gen-Llama3_1-70BI1-IQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
New-Dawn-Llama-3-70B-32K-v1.0I1-IQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
Meta-Llama-3-70B-Instruct-abliterated-v3.5I1-IQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
Athene-70BIQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
L3.3-70B-Magnum-DiamondIQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 GiB14±8.3%
Meta-Llama-3-70B-InstructIQ3_XXS70.6B25.58 GiB21.25 GiB47.51 GiB0.49 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?
1988 of 2118 indexed open-weight models fit a Apple M2 Ultra at 131,072 context with q8_0 KV cache, the largest being Apertus-70B-Instruct-2509 at UD-IQ3_XXS. 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.