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 f16 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 GiB2.14 GiB383.90 GiB0.10 GiB10±37%
DeepSeek-V3.1-TerminusMoEQ4_K_M685B381.12 GiB2.14 GiB383.90 GiB0.10 GiB10±37%
DeepSeek-V3.1MoEQ4_K_M685B381.12 GiB2.14 GiB383.90 GiB0.10 GiB10±37%
Step-3.7-FlashBF16201B366.96 GiB13.03 GiB380.57 GiB3.43 GiB2±8.3%
DeepSeek-V3.2MoEQ4_K_M685B377.56 GiB2.14 GiB380.33 GiB3.67 GiB11±37%
cogito-671b-v2.1MoEQ4_K_M671B377.55 GiB2.14 GiB380.33 GiB3.67 GiB11±37%
Kimi-K2.5MoEIQ3_S1059B377.51 GiB2.14 GiB380.28 GiB3.72 GiB12±37%
DeepSeek-TNG-R1T2-ChimeraMoEQ4_K_M685B377.13 GiB2.14 GiB379.90 GiB4.10 GiB11±37%
cogito-v2-preview-deepseek-671B-MoEMoEQ4_K_M671B377.13 GiB2.14 GiB379.90 GiB4.10 GiB11±37%
DeepSeek-V3-0324MoEQ4_K_M685B376.89 GiB2.14 GiB379.67 GiB4.33 GiB11±37%
DeepSeek-Prover-V2-671BMoEQ4_K_M685B376.71 GiB2.14 GiB379.49 GiB4.51 GiB11±37%
r1-1776MoEQ4_K_M671B376.65 GiB2.14 GiB379.43 GiB4.57 GiB11±37%
DeepSeek-R1MoEQ4_K_M685B376.65 GiB2.14 GiB379.43 GiB4.57 GiB11±37%
GLM-5.1MoEIQ4_XS754B375.69 GiB2.74 GiB379.03 GiB4.97 GiB10±37%
GLM-5MoEIQ4_XS754B375.23 GiB2.74 GiB378.57 GiB5.43 GiB10±37%
Qwen3-Coder-480B-A35B-InstructMoEQ6_K480B367.10 GiB7.75 GiB375.44 GiB8.56 GiB8±37%
Kimi-K2-ThinkingMoEIQ3_XXS1058B367.09 GiB2.14 GiB369.86 GiB14.14 GiB12±37%
Qwen3-Coder-REAP-363B-A35BMoEQ8_0363B359.45 GiB7.75 GiB367.78 GiB16.22 GiB7±37%
GLM-4.5MoEQ8_0358B354.80 GiB11.50 GiB366.89 GiB17.11 GiB8±37%
GLM-4.7MoEQ8_0358B354.80 GiB11.50 GiB366.89 GiB17.11 GiB8±37%
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16MoEUD-Q5_K_S561B364.80 GiB0.00 GiB365.38 GiB18.62 GiB10±37%
GLM-4.6MoEQ8_0357B353.27 GiB11.50 GiB365.36 GiB18.64 GiB8±37%
GLM-4.6-Derestricted-v3MoEQ8_0357B353.27 GiB11.50 GiB365.36 GiB18.64 GiB8±37%
MiMo-V2.5-ProMoEKV unresolvedUD-IQ3_S1023B351.94 GiB10.94 GiB363.49 GiB20.51 GiB10±37%
Kimi-K2.7-CodeMoEUD-IQ3_XXS1059B351.00 GiB2.14 GiB353.77 GiB30.23 GiB12±37%
Kimi-K2-InstructMoEQ2_K_L1026B347.81 GiB2.14 GiB350.58 GiB33.42 GiB12±37%
GLM-5.2MoEUD-IQ4_NL753B347.07 GiB2.74 GiB350.41 GiB33.59 GiB11±37%
MiniMax-M3MoEQ6_K427B344.03 GiB3.75 GiB348.35 GiB35.65 GiB11±37%
Kimi-K2.6MoEQ2_K_L1059B334.66 GiB2.14 GiB337.43 GiB46.57 GiB13±37%
Qwen3.5-397B-A17BMoEQ6_K_L403B325.38 GiB0.94 GiB326.91 GiB57.09 GiB14±37%
Hermes-3-Llama-3.1-405BQ6_K406B310.08 GiB15.75 GiB326.66 GiB57.34 GiB2±8.3%
Hermes-4-405BQ6_K406B310.08 GiB15.75 GiB326.66 GiB57.34 GiB2±8.3%
Trinity-Large-ThinkingMoEQ6_K_L399B319.94 GiB2.67 GiB323.18 GiB60.82 GiB14±37%
Ornith-1.0-397BMoEQ6_K397B318.87 GiB0.94 GiB320.41 GiB63.59 GiB14±37%
Nex-N2-ProMoEQ6_K397B318.87 GiB0.94 GiB320.41 GiB63.59 GiB14±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ6_K402B306.19 GiB6.00 GiB312.77 GiB71.23 GiB14±37%
MiMo-V2.5MoEKV unresolvedQ8_0311B306.67 GiB3.75 GiB311.01 GiB72.99 GiB12±37%
MiMo-V2-FlashMoEKV unresolvedQ8_0310B305.69 GiB3.75 GiB310.04 GiB73.96 GiB12±37%
Trinity-Large-PreviewMoEQ6_K_L399B305.35 GiB2.67 GiB308.59 GiB75.41 GiB14±37%
Trinity-Large-TrueBaseMoEQ6_K399B305.07 GiB2.67 GiB308.31 GiB75.69 GiB14±37%
Hy3MoEQ8_0299B295.84 GiB10.00 GiB306.42 GiB77.58 GiB10±37%
ERNIE-4.5-300B-A47B-PTQ8_0300B296.43 GiB6.75 GiB303.85 GiB80.15 GiB2±8.3%
dots.llm1.instMoEBF16143B266.00 GiB31.00 GiB297.58 GiB86.42 GiB7±37%
GLM-4.6-REAP-268B-A32BMoEQ8_0269B266.14 GiB11.50 GiB278.23 GiB105.77 GiB9±37%
grok-2MoEQ8_0270B266.72 GiB8.00 GiB275.41 GiB108.59 GiB5±37%
InklingMoEUD-IQ1_M952B265.46 GiB8.25 GiB274.28 GiB109.72 GiB12±37%
Athene-70BF3270.6B262.84 GiB10.00 GiB273.52 GiB110.48 GiB3±8.3%
Solar-Open2-250BMoEQ8_0250B247.86 GiB6.00 GiB254.44 GiB129.56 GiB14±37%
Devstral-2-123B-Instruct-2512BF16125B232.89 GiB11.00 GiB244.60 GiB139.40 GiB3±8.3%
Mistral-Medium-3.5-128BBF16128B232.89 GiB11.00 GiB244.60 GiB139.40 GiB3±8.3%
Qwen3-VL-235B-A22B-InstructMoEQ8_0236B232.77 GiB5.88 GiB239.23 GiB144.77 GiB11±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ8_0236B232.77 GiB5.88 GiB239.23 GiB144.77 GiB11±37%
Qwen3-235B-A22BMoEQ8_0235B232.77 GiB5.88 GiB239.23 GiB144.77 GiB11±37%
Qwen3-235B-A22B-Thinking-2507MoEQ8_0235B232.77 GiB5.88 GiB239.23 GiB144.77 GiB11±37%
Qwen3-235B-A22B-Instruct-2507MoEQ8_0235B232.77 GiB5.88 GiB239.23 GiB144.77 GiB11±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ8_0236B233.41 GiB2.11 GiB236.11 GiB147.89 GiB15±37%
DeepSeek-V2.5MoEQ8_0236B233.41 GiB2.11 GiB236.11 GiB147.89 GiB15±37%
DeepSeek-Coder-V2-InstructMoEQ8_0236B233.41 GiB2.11 GiB236.11 GiB147.89 GiB15±37%
MiniMax-M2.1MoEQ8_0229B226.44 GiB7.75 GiB234.72 GiB149.28 GiB13±37%
MiniMax-M2.5MoEQ8_0229B226.44 GiB7.75 GiB234.72 GiB149.28 GiB13±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 f16 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.