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 16K context with q8_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 16K 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 GiB0.57 GiB382.32 GiB1.68 GiB11±37%
DeepSeek-V3.1-TerminusMoEQ4_K_M685B381.12 GiB0.57 GiB382.32 GiB1.68 GiB11±37%
DeepSeek-V3.1MoEQ4_K_M685B381.12 GiB0.57 GiB382.32 GiB1.68 GiB11±37%
DeepSeek-V3.2MoEQ4_K_M685B377.56 GiB0.57 GiB378.75 GiB5.25 GiB11±37%
cogito-671b-v2.1MoEQ4_K_M671B377.55 GiB0.57 GiB378.75 GiB5.25 GiB11±37%
Kimi-K2.5MoEIQ3_S1059B377.51 GiB0.57 GiB378.71 GiB5.29 GiB12±37%
DeepSeek-TNG-R1T2-ChimeraMoEQ4_K_M685B377.13 GiB0.57 GiB378.33 GiB5.67 GiB11±37%
cogito-v2-preview-deepseek-671B-MoEMoEQ4_K_M671B377.13 GiB0.57 GiB378.33 GiB5.67 GiB11±37%
DeepSeek-V3-0324MoEQ4_K_M685B376.89 GiB0.57 GiB378.09 GiB5.91 GiB11±37%
DeepSeek-Prover-V2-671BMoEQ4_K_M685B376.71 GiB0.57 GiB377.91 GiB6.09 GiB11±37%
r1-1776MoEQ4_K_M671B376.65 GiB0.57 GiB377.85 GiB6.15 GiB11±37%
DeepSeek-R1MoEQ4_K_M685B376.65 GiB0.57 GiB377.85 GiB6.15 GiB11±37%
GLM-5.1MoEIQ4_XS754B375.69 GiB0.73 GiB377.02 GiB6.98 GiB11±37%
GLM-5MoEIQ4_XS754B375.23 GiB0.73 GiB376.55 GiB7.45 GiB11±37%
Step-3.7-FlashBF16201B366.96 GiB3.74 GiB371.27 GiB12.73 GiB2±8.3%
Qwen3-Coder-480B-A35B-InstructMoEQ6_K480B367.10 GiB2.06 GiB369.74 GiB14.26 GiB9±37%
Kimi-K2-ThinkingMoEIQ3_XXS1058B367.09 GiB0.57 GiB368.29 GiB15.71 GiB13±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 GiB2.06 GiB362.09 GiB21.91 GiB8±37%
GLM-4.5MoEQ8_0358B354.80 GiB3.05 GiB358.45 GiB25.55 GiB9±37%
GLM-4.7MoEQ8_0358B354.80 GiB3.05 GiB358.45 GiB25.55 GiB9±37%
GLM-4.6MoEQ8_0357B353.27 GiB3.05 GiB356.91 GiB27.09 GiB9±37%
GLM-4.6-Derestricted-v3MoEQ8_0357B353.27 GiB3.05 GiB356.91 GiB27.09 GiB9±37%
MiMo-V2.5-ProMoEKV unresolvedUD-IQ3_S1023B351.94 GiB2.91 GiB355.46 GiB28.54 GiB12±37%
Kimi-K2.7-CodeMoEUD-IQ3_XXS1059B351.00 GiB0.57 GiB352.20 GiB31.80 GiB13±37%
Kimi-K2-InstructMoEQ2_K_L1026B347.81 GiB0.57 GiB349.01 GiB34.99 GiB13±37%
GLM-5.2MoEUD-IQ4_NL753B347.07 GiB0.73 GiB348.39 GiB35.61 GiB12±37%
MiniMax-M3MoEQ6_K427B344.03 GiB1.00 GiB345.59 GiB38.41 GiB12±37%
Kimi-K2.6MoEQ2_K_L1059B334.66 GiB0.57 GiB335.86 GiB48.14 GiB14±37%
Qwen3.5-397B-A17BMoEQ6_K_L403B325.38 GiB0.25 GiB326.23 GiB57.77 GiB14±37%
Trinity-Large-ThinkingMoEQ6_K_L399B319.94 GiB0.92 GiB321.44 GiB62.56 GiB15±37%
Ornith-1.0-397BMoEQ6_K397B318.87 GiB0.25 GiB319.72 GiB64.28 GiB15±37%
Nex-N2-ProMoEQ6_K397B318.87 GiB0.25 GiB319.72 GiB64.28 GiB15±37%
Hermes-3-Llama-3.1-405BQ6_K406B310.08 GiB4.18 GiB315.10 GiB68.90 GiB2±8.3%
Hermes-4-405BQ6_K406B310.08 GiB4.18 GiB315.10 GiB68.90 GiB2±8.3%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ6_K402B306.19 GiB1.59 GiB308.36 GiB75.64 GiB16±37%
MiMo-V2.5MoEKV unresolvedQ8_0311B306.67 GiB1.00 GiB308.26 GiB75.74 GiB13±37%
MiMo-V2-FlashMoEKV unresolvedQ8_0310B305.69 GiB1.00 GiB307.28 GiB76.72 GiB13±37%
Trinity-Large-PreviewMoEQ6_K_L399B305.35 GiB0.92 GiB306.84 GiB77.16 GiB15±37%
Trinity-Large-TrueBaseMoEQ6_K399B305.07 GiB0.92 GiB306.56 GiB77.44 GiB15±37%
Hy3MoEQ8_0299B295.84 GiB2.66 GiB299.08 GiB84.92 GiB12±37%
ERNIE-4.5-300B-A47B-PTQ8_0300B296.43 GiB1.79 GiB298.89 GiB85.11 GiB2±8.3%
dots.llm1.instMoEBF16143B266.00 GiB8.23 GiB274.82 GiB109.18 GiB11±37%
GLM-4.6-REAP-268B-A32BMoEQ8_0269B266.14 GiB3.05 GiB269.78 GiB114.22 GiB11±37%
grok-2MoEQ8_0270B266.72 GiB2.13 GiB269.53 GiB114.47 GiB5±37%
InklingMoEUD-IQ1_M952B265.46 GiB2.19 GiB268.22 GiB115.78 GiB15±37%
Athene-70BF3270.6B262.84 GiB2.66 GiB266.17 GiB117.83 GiB3±8.3%
Solar-Open2-250BMoEQ8_0250B247.86 GiB1.59 GiB250.03 GiB133.97 GiB16±37%
Devstral-2-123B-Instruct-2512BF16125B232.89 GiB2.92 GiB236.52 GiB147.48 GiB3±8.3%
Mistral-Medium-3.5-128BBF16128B232.89 GiB2.92 GiB236.52 GiB147.48 GiB3±8.3%
Qwen3-VL-235B-A22B-InstructMoEQ8_0236B232.77 GiB1.56 GiB234.92 GiB149.08 GiB13±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ8_0236B232.77 GiB1.56 GiB234.92 GiB149.08 GiB13±37%
Qwen3-235B-A22BMoEQ8_0235B232.77 GiB1.56 GiB234.92 GiB149.08 GiB13±37%
Qwen3-235B-A22B-Thinking-2507MoEQ8_0235B232.77 GiB1.56 GiB234.92 GiB149.08 GiB13±37%
Qwen3-235B-A22B-Instruct-2507MoEQ8_0235B232.77 GiB1.56 GiB234.92 GiB149.08 GiB13±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ8_0236B233.41 GiB0.56 GiB234.56 GiB149.44 GiB16±37%
DeepSeek-V2.5MoEQ8_0236B233.41 GiB0.56 GiB234.56 GiB149.44 GiB16±37%
DeepSeek-Coder-V2-InstructMoEQ8_0236B233.41 GiB0.56 GiB234.56 GiB149.44 GiB16±37%
Qwen3.5-122B-A10B-hereticMoEBF16123B232.24 GiB0.20 GiB233.01 GiB150.99 GiB17±37%
Qwen3.5-122B-A10BMoEBF16125B232.24 GiB0.20 GiB233.01 GiB150.99 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 16,384 context with q8_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.