Apple · apple

Apple M3 Ultra

Apple M3 Ultra has 512 GB of unified memory at 819 GB/s — about 357.12 GiB usable after driver and compositor overhead. 2116 of 2118 indexed models fit at 32K context with q4_0 KV. Note only 384 GB of its 512 GB is allocatable to the GPU.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
512 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 1817vision language 195image 2audio tts 21audio asr 39video 16embedding 26

What fits at 32K context

largest quantization that fits, per model · 2116 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
DeepSeek-R1-0528MoEQ4_K_M685B381.12 GiB0.60 GiB382.36 GiB1.64 GiB11±37%
DeepSeek-V3.1-TerminusMoEQ4_K_M685B381.12 GiB0.60 GiB382.36 GiB1.64 GiB11±37%
DeepSeek-V3.1MoEQ4_K_M685B381.12 GiB0.60 GiB382.36 GiB1.64 GiB11±37%
DeepSeek-V3.2MoEQ4_K_M685B377.56 GiB0.60 GiB378.79 GiB5.21 GiB11±37%
cogito-671b-v2.1MoEQ4_K_M671B377.55 GiB0.60 GiB378.78 GiB5.22 GiB11±37%
Kimi-K2.5MoEIQ3_S1059B377.51 GiB0.60 GiB378.74 GiB5.26 GiB12±37%
DeepSeek-TNG-R1T2-ChimeraMoEQ4_K_M685B377.13 GiB0.60 GiB378.36 GiB5.64 GiB11±37%
cogito-v2-preview-deepseek-671B-MoEMoEQ4_K_M671B377.13 GiB0.60 GiB378.36 GiB5.64 GiB11±37%
DeepSeek-V3-0324MoEQ4_K_M685B376.89 GiB0.60 GiB378.13 GiB5.87 GiB11±37%
DeepSeek-Prover-V2-671BMoEQ4_K_M685B376.71 GiB0.60 GiB377.95 GiB6.05 GiB11±37%
r1-1776MoEQ4_K_M671B376.65 GiB0.60 GiB377.89 GiB6.11 GiB11±37%
DeepSeek-R1MoEQ4_K_M685B376.65 GiB0.60 GiB377.89 GiB6.11 GiB11±37%
GLM-5.1MoEIQ4_XS754B375.69 GiB0.77 GiB377.06 GiB6.94 GiB11±37%
GLM-5MoEIQ4_XS754B375.23 GiB0.77 GiB376.60 GiB7.40 GiB11±37%
Step-3.7-FlashBF16201B366.96 GiB3.67 GiB371.20 GiB12.80 GiB2±8.3%
Qwen3-Coder-480B-A35B-InstructMoEQ6_K480B367.10 GiB2.18 GiB369.87 GiB14.13 GiB9±37%
Kimi-K2-ThinkingMoEIQ3_XXS1058B367.09 GiB0.60 GiB368.32 GiB15.68 GiB12±37%
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16MoEUD-Q5_K_S561B364.80 GiB0.00 GiB365.38 GiB18.62 GiB10±37%
Qwen3-Coder-REAP-363B-A35BMoEQ8_0363B359.45 GiB2.18 GiB362.21 GiB21.79 GiB8±37%
GLM-4.5MoEQ8_0358B354.80 GiB3.23 GiB358.63 GiB25.37 GiB9±37%
GLM-4.7MoEQ8_0358B354.80 GiB3.23 GiB358.63 GiB25.37 GiB9±37%
GLM-4.6MoEQ8_0357B353.27 GiB3.23 GiB357.09 GiB26.91 GiB9±37%
GLM-4.6-Derestricted-v3MoEQ8_0357B353.27 GiB3.23 GiB357.09 GiB26.91 GiB9±37%
MiMo-V2.5-ProMoEKV unresolvedUD-IQ3_S1023B351.94 GiB3.08 GiB355.63 GiB28.37 GiB12±37%
Kimi-K2.7-CodeMoEUD-IQ3_XXS1059B351.00 GiB0.60 GiB352.23 GiB31.77 GiB13±37%
Kimi-K2-InstructMoEQ2_K_L1026B347.81 GiB0.60 GiB349.04 GiB34.96 GiB13±37%
GLM-5.2MoEUD-IQ4_NL753B347.07 GiB0.77 GiB348.44 GiB35.56 GiB12±37%
MiniMax-M3MoEQ6_K427B344.03 GiB1.05 GiB345.65 GiB38.35 GiB12±37%
Kimi-K2.6MoEQ2_K_L1059B334.66 GiB0.60 GiB335.89 GiB48.11 GiB14±37%
Qwen3.5-397B-A17BMoEQ6_K_L403B325.38 GiB0.26 GiB326.24 GiB57.76 GiB14±37%
Trinity-Large-ThinkingMoEQ6_K_L399B319.94 GiB0.75 GiB321.27 GiB62.73 GiB15±37%
Ornith-1.0-397BMoEQ6_K397B318.87 GiB0.26 GiB319.74 GiB64.26 GiB15±37%
Nex-N2-ProMoEQ6_K397B318.87 GiB0.26 GiB319.74 GiB64.26 GiB15±37%
Hermes-3-Llama-3.1-405BQ6_K406B310.08 GiB4.43 GiB315.34 GiB68.66 GiB2±8.3%
Hermes-4-405BQ6_K406B310.08 GiB4.43 GiB315.34 GiB68.66 GiB2±8.3%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ6_K402B306.19 GiB1.69 GiB308.46 GiB75.54 GiB16±37%
MiMo-V2.5MoEKV unresolvedQ8_0311B306.67 GiB1.05 GiB308.32 GiB75.68 GiB13±37%
MiMo-V2-FlashMoEKV unresolvedQ8_0310B305.69 GiB1.05 GiB307.34 GiB76.66 GiB13±37%
Trinity-Large-PreviewMoEQ6_K_L399B305.35 GiB0.75 GiB306.67 GiB77.33 GiB15±37%
Trinity-Large-TrueBaseMoEQ6_K399B305.07 GiB0.75 GiB306.40 GiB77.60 GiB16±37%
Hy3MoEQ8_0299B295.84 GiB2.81 GiB299.23 GiB84.77 GiB12±37%
ERNIE-4.5-300B-A47B-PTQ8_0300B296.43 GiB1.90 GiB299.00 GiB85.00 GiB2±8.3%
dots.llm1.instMoEBF16143B266.00 GiB8.72 GiB275.30 GiB108.70 GiB10±37%
GLM-4.6-REAP-268B-A32BMoEQ8_0269B266.14 GiB3.23 GiB269.96 GiB114.04 GiB11±37%
grok-2MoEQ8_0270B266.72 GiB2.25 GiB269.66 GiB114.34 GiB5±37%
InklingMoEUD-IQ1_M952B265.46 GiB2.32 GiB268.35 GiB115.65 GiB15±37%
Athene-70BF3270.6B262.84 GiB2.81 GiB266.33 GiB117.67 GiB3±8.3%
Solar-Open2-250BMoEQ8_0250B247.86 GiB1.69 GiB250.12 GiB133.88 GiB16±37%
Devstral-2-123B-Instruct-2512BF16125B232.89 GiB3.09 GiB236.69 GiB147.31 GiB3±8.3%
Mistral-Medium-3.5-128BBF16128B232.89 GiB3.09 GiB236.69 GiB147.31 GiB3±8.3%
Qwen3-VL-235B-A22B-InstructMoEQ8_0236B232.77 GiB1.65 GiB235.01 GiB148.99 GiB13±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ8_0236B232.77 GiB1.65 GiB235.01 GiB148.99 GiB13±37%
Qwen3-235B-A22BMoEQ8_0235B232.77 GiB1.65 GiB235.01 GiB148.99 GiB13±37%
Qwen3-235B-A22B-Thinking-2507MoEQ8_0235B232.77 GiB1.65 GiB235.01 GiB148.99 GiB13±37%
Qwen3-235B-A22B-Instruct-2507MoEQ8_0235B232.77 GiB1.65 GiB235.01 GiB148.99 GiB13±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ8_0236B233.41 GiB0.59 GiB234.59 GiB149.41 GiB16±37%
DeepSeek-V2.5MoEQ8_0236B233.41 GiB0.59 GiB234.59 GiB149.41 GiB16±37%
DeepSeek-Coder-V2-InstructMoEQ8_0236B233.41 GiB0.59 GiB234.59 GiB149.41 GiB16±37%
Qwen3.5-122B-A10B-hereticMoEBF16123B232.24 GiB0.21 GiB233.03 GiB150.97 GiB17±37%
Qwen3.5-122B-A10BMoEBF16125B232.24 GiB0.21 GiB233.02 GiB150.98 GiB17±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 M3 Ultra run?
2116 of 2118 indexed open-weight models fit a Apple M3 Ultra at 32,768 context with q4_0 KV cache, the largest being DeepSeek-R1-0528 at Q4_K_M. That covers text, vision-language, image, video and speech models.
How much usable memory does a Apple M3 Ultra actually have?
Its nameplate is 512 GB, but about 357.12 GiB is available to a model once driver and compositor overhead is accounted for, and only 384 GB of the pool can be allocated to the GPU at all.
Is a Apple M3 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.