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 8K 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 8K 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 GiB0.28 GiB382.04 GiB1.96 GiB11±37%
DeepSeek-V3.1-TerminusMoEQ4_K_M685B381.12 GiB0.28 GiB382.04 GiB1.96 GiB11±37%
DeepSeek-V3.1MoEQ4_K_M685B381.12 GiB0.28 GiB382.04 GiB1.96 GiB11±37%
DeepSeek-V3.2MoEQ4_K_M685B377.56 GiB0.28 GiB378.47 GiB5.53 GiB11±37%
cogito-671b-v2.1MoEQ4_K_M671B377.55 GiB0.28 GiB378.47 GiB5.53 GiB11±37%
Kimi-K2.5MoEIQ3_S1059B377.51 GiB0.28 GiB378.42 GiB5.58 GiB12±37%
DeepSeek-TNG-R1T2-ChimeraMoEQ4_K_M685B377.13 GiB0.28 GiB378.05 GiB5.95 GiB11±37%
cogito-v2-preview-deepseek-671B-MoEMoEQ4_K_M671B377.13 GiB0.28 GiB378.04 GiB5.96 GiB11±37%
DeepSeek-V3-0324MoEQ4_K_M685B376.89 GiB0.28 GiB377.81 GiB6.19 GiB11±37%
DeepSeek-Prover-V2-671BMoEQ4_K_M685B376.71 GiB0.28 GiB377.63 GiB6.37 GiB11±37%
r1-1776MoEQ4_K_M671B376.65 GiB0.28 GiB377.57 GiB6.43 GiB11±37%
DeepSeek-R1MoEQ4_K_M685B376.65 GiB0.28 GiB377.57 GiB6.43 GiB11±37%
GLM-5.1MoEIQ4_XS754B375.69 GiB0.36 GiB376.65 GiB7.35 GiB11±37%
GLM-5MoEIQ4_XS754B375.23 GiB0.36 GiB376.19 GiB7.81 GiB11±37%
Step-3.7-FlashBF16201B366.96 GiB2.14 GiB369.68 GiB14.32 GiB2±8.3%
Qwen3-Coder-480B-A35B-InstructMoEQ6_K480B367.10 GiB1.03 GiB368.71 GiB15.29 GiB9±37%
Kimi-K2-ThinkingMoEIQ3_XXS1058B367.09 GiB0.28 GiB368.00 GiB16.00 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 GiB1.03 GiB361.06 GiB22.94 GiB8±37%
GLM-4.5MoEQ8_0358B354.80 GiB1.53 GiB356.92 GiB27.08 GiB10±37%
GLM-4.7MoEQ8_0358B354.80 GiB1.53 GiB356.92 GiB27.08 GiB10±37%
GLM-4.6MoEQ8_0357B353.27 GiB1.53 GiB355.39 GiB28.61 GiB10±37%
GLM-4.6-Derestricted-v3MoEQ8_0357B353.27 GiB1.53 GiB355.39 GiB28.61 GiB10±37%
MiMo-V2.5-ProMoEKV unresolvedUD-IQ3_S1023B351.94 GiB1.45 GiB354.01 GiB29.99 GiB13±37%
Kimi-K2.7-CodeMoEUD-IQ3_XXS1059B351.00 GiB0.28 GiB351.91 GiB32.09 GiB13±37%
Kimi-K2-InstructMoEQ2_K_L1026B347.81 GiB0.28 GiB348.72 GiB35.28 GiB13±37%
GLM-5.2MoEUD-IQ4_NL753B347.07 GiB0.36 GiB348.03 GiB35.97 GiB12±37%
MiniMax-M3MoEQ6_K427B344.03 GiB0.50 GiB345.09 GiB38.91 GiB12±37%
Kimi-K2.6MoEQ2_K_L1059B334.66 GiB0.28 GiB335.57 GiB48.43 GiB14±37%
Qwen3.5-397B-A17BMoEQ6_K_L403B325.38 GiB0.12 GiB326.10 GiB57.90 GiB14±37%
Trinity-Large-ThinkingMoEQ6_K_L399B319.94 GiB0.67 GiB321.19 GiB62.81 GiB15±37%
Ornith-1.0-397BMoEQ6_K397B318.87 GiB0.12 GiB319.60 GiB64.40 GiB15±37%
Nex-N2-ProMoEQ6_K397B318.87 GiB0.12 GiB319.60 GiB64.40 GiB15±37%
Hermes-3-Llama-3.1-405BQ6_K406B310.08 GiB2.09 GiB313.01 GiB70.99 GiB2±8.3%
Hermes-4-405BQ6_K406B310.08 GiB2.09 GiB313.01 GiB70.99 GiB2±8.3%
MiMo-V2.5MoEKV unresolvedQ8_0311B306.67 GiB0.50 GiB307.76 GiB76.24 GiB13±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ6_K402B306.19 GiB0.80 GiB307.57 GiB76.43 GiB17±37%
MiMo-V2-FlashMoEKV unresolvedQ8_0310B305.69 GiB0.50 GiB306.78 GiB77.22 GiB13±37%
Trinity-Large-PreviewMoEQ6_K_L399B305.35 GiB0.67 GiB306.59 GiB77.41 GiB16±37%
Trinity-Large-TrueBaseMoEQ6_K399B305.07 GiB0.67 GiB306.32 GiB77.68 GiB16±37%
ERNIE-4.5-300B-A47B-PTQ8_0300B296.43 GiB0.90 GiB298.00 GiB86.00 GiB2±8.3%
Hy3MoEQ8_0299B295.84 GiB1.33 GiB297.75 GiB86.25 GiB12±37%
dots.llm1.instMoEBF16143B266.00 GiB4.12 GiB270.70 GiB113.30 GiB12±37%
grok-2MoEQ8_0270B266.72 GiB1.06 GiB268.47 GiB115.53 GiB5±37%
GLM-4.6-REAP-268B-A32BMoEQ8_0269B266.14 GiB1.53 GiB268.26 GiB115.74 GiB11±37%
InklingMoEUD-IQ1_M952B265.46 GiB1.10 GiB267.12 GiB116.88 GiB16±37%
Athene-70BF3270.6B262.84 GiB1.33 GiB264.84 GiB119.16 GiB3±8.3%
Solar-Open2-250BMoEQ8_0250B247.86 GiB0.80 GiB249.23 GiB134.77 GiB17±37%
Devstral-2-123B-Instruct-2512BF16125B232.89 GiB1.46 GiB235.06 GiB148.94 GiB3±8.3%
Mistral-Medium-3.5-128BBF16128B232.89 GiB1.46 GiB235.06 GiB148.94 GiB3±8.3%
DeepSeek-Coder-V2-Instruct-0724MoEQ8_0236B233.41 GiB0.28 GiB234.28 GiB149.72 GiB16±37%
DeepSeek-V2.5MoEQ8_0236B233.41 GiB0.28 GiB234.28 GiB149.72 GiB16±37%
DeepSeek-Coder-V2-InstructMoEQ8_0236B233.41 GiB0.28 GiB234.28 GiB149.72 GiB16±37%
Qwen3-VL-235B-A22B-InstructMoEQ8_0236B232.77 GiB0.78 GiB234.14 GiB149.86 GiB13±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ8_0236B232.77 GiB0.78 GiB234.14 GiB149.86 GiB13±37%
Qwen3-235B-A22BMoEQ8_0235B232.77 GiB0.78 GiB234.14 GiB149.86 GiB13±37%
Qwen3-235B-A22B-Thinking-2507MoEQ8_0235B232.77 GiB0.78 GiB234.14 GiB149.86 GiB13±37%
Qwen3-235B-A22B-Instruct-2507MoEQ8_0235B232.77 GiB0.78 GiB234.14 GiB149.86 GiB13±37%
Qwen3.5-122B-A10B-hereticMoEBF16123B232.24 GiB0.10 GiB232.91 GiB151.09 GiB17±37%
Qwen3.5-122B-A10BMoEBF16125B232.24 GiB0.10 GiB232.91 GiB151.09 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 8,192 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.