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 128K 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
vision language 195text 1817image 2audio tts 21audio asr 39video 16embedding 26

What fits at 128K context

largest quantization that fits, per model · 2116 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Step-3.7-FlashBF16201B366.96 GiB13.79 GiB381.33 GiB2.67 GiB2±8.3%
DeepSeek-V3.2MoEQ4_K_M685B377.56 GiB2.41 GiB380.60 GiB3.40 GiB10±37%
cogito-671b-v2.1MoEQ4_K_M671B377.55 GiB2.41 GiB380.59 GiB3.41 GiB10±37%
Kimi-K2.5MoEIQ3_S1059B377.51 GiB2.41 GiB380.55 GiB3.45 GiB12±37%
DeepSeek-TNG-R1T2-ChimeraMoEQ4_K_M685B377.13 GiB2.41 GiB380.17 GiB3.83 GiB10±37%
cogito-v2-preview-deepseek-671B-MoEMoEQ4_K_M671B377.13 GiB2.41 GiB380.17 GiB3.83 GiB10±37%
DeepSeek-V3-0324MoEQ4_K_M685B376.89 GiB2.41 GiB379.93 GiB4.07 GiB10±37%
DeepSeek-Prover-V2-671BMoEQ4_K_M685B376.71 GiB2.41 GiB379.76 GiB4.24 GiB10±37%
r1-1776MoEQ4_K_M671B376.65 GiB2.41 GiB379.70 GiB4.30 GiB10±37%
DeepSeek-R1MoEQ4_K_M685B376.65 GiB2.41 GiB379.70 GiB4.30 GiB10±37%
GLM-5.1MoEIQ4_XS754B375.69 GiB3.08 GiB379.37 GiB4.63 GiB10±37%
GLM-5MoEIQ4_XS754B375.23 GiB3.08 GiB378.91 GiB5.09 GiB10±37%
Qwen3-Coder-480B-A35B-InstructMoEQ6_K480B367.10 GiB8.72 GiB376.40 GiB7.60 GiB8±37%
Kimi-K2-ThinkingMoEIQ3_XXS1058B367.09 GiB2.41 GiB370.13 GiB13.87 GiB12±37%
DeepSeek-R1-0528MoEQ4_K_S685B367.08 GiB2.41 GiB370.13 GiB13.87 GiB11±37%
DeepSeek-V3.1-TerminusMoEQ4_K_S685B367.08 GiB2.41 GiB370.13 GiB13.87 GiB11±37%
DeepSeek-V3.1MoEQ4_K_S685B367.08 GiB2.41 GiB370.13 GiB13.87 GiB11±37%
Qwen3-Coder-REAP-363B-A35BMoEQ8_0363B359.45 GiB8.72 GiB368.75 GiB15.25 GiB7±37%
GLM-4.5MoEQ8_0358B354.80 GiB12.94 GiB368.33 GiB15.67 GiB8±37%
GLM-4.7MoEQ8_0358B354.80 GiB12.94 GiB368.33 GiB15.67 GiB8±37%
GLM-4.6MoEQ8_0357B353.27 GiB12.94 GiB366.80 GiB17.20 GiB8±37%
GLM-4.6-Derestricted-v3MoEQ8_0357B353.27 GiB12.94 GiB366.80 GiB17.20 GiB8±37%
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16MoEUD-Q5_K_S561B364.80 GiB0.00 GiB365.38 GiB18.62 GiB10±37%
MiMo-V2.5-ProMoEKV unresolvedUD-IQ3_S1023B351.94 GiB12.30 GiB364.86 GiB19.14 GiB9±37%
Kimi-K2.7-CodeMoEUD-IQ3_XXS1059B351.00 GiB2.41 GiB354.04 GiB29.96 GiB12±37%
Kimi-K2-InstructMoEQ2_K_L1026B347.81 GiB2.41 GiB350.85 GiB33.15 GiB12±37%
GLM-5.2MoEUD-IQ4_NL753B347.07 GiB3.08 GiB350.75 GiB33.25 GiB11±37%
MiniMax-M3MoEQ6_K427B344.03 GiB4.22 GiB348.81 GiB35.19 GiB11±37%
Kimi-K2.6MoEQ2_K_L1059B334.66 GiB2.41 GiB337.70 GiB46.30 GiB13±37%
Hermes-3-Llama-3.1-405BQ6_K406B310.08 GiB17.72 GiB328.63 GiB55.37 GiB2±8.3%
Hermes-4-405BQ6_K406B310.08 GiB17.72 GiB328.63 GiB55.37 GiB2±8.3%
Qwen3.5-397B-A17BMoEQ6_K_L403B325.38 GiB1.05 GiB327.03 GiB56.97 GiB14±37%
Trinity-Large-ThinkingMoEQ6_K_L399B319.94 GiB2.33 GiB322.85 GiB61.15 GiB14±37%
Ornith-1.0-397BMoEQ6_K397B318.87 GiB1.05 GiB320.53 GiB63.47 GiB14±37%
Nex-N2-ProMoEQ6_K397B318.87 GiB1.05 GiB320.53 GiB63.47 GiB14±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ6_K402B306.19 GiB6.75 GiB313.52 GiB70.48 GiB14±37%
MiMo-V2.5MoEKV unresolvedQ8_0311B306.67 GiB4.22 GiB311.48 GiB72.52 GiB12±37%
MiMo-V2-FlashMoEKV unresolvedQ8_0310B305.69 GiB4.22 GiB310.50 GiB73.50 GiB12±37%
Trinity-Large-PreviewMoEQ6_K_L399B305.35 GiB2.33 GiB308.26 GiB75.74 GiB15±37%
Trinity-Large-TrueBaseMoEQ6_K399B305.07 GiB2.33 GiB307.98 GiB76.02 GiB15±37%
Hy3MoEQ8_0299B295.84 GiB11.25 GiB307.67 GiB76.33 GiB9±37%
ERNIE-4.5-300B-A47B-PTQ8_0300B296.43 GiB7.59 GiB304.70 GiB79.30 GiB2±8.3%
dots.llm1.instMoEBF16143B266.00 GiB34.88 GiB301.46 GiB82.54 GiB6±37%
GLM-4.6-REAP-268B-A32BMoEQ8_0269B266.14 GiB12.94 GiB279.67 GiB104.33 GiB8±37%
grok-2MoEQ8_0270B266.72 GiB9.00 GiB276.41 GiB107.59 GiB4±37%
InklingMoEUD-IQ1_M952B265.46 GiB9.28 GiB275.31 GiB108.69 GiB12±37%
Athene-70BF3270.6B262.84 GiB11.25 GiB274.77 GiB109.23 GiB3±8.3%
Solar-Open2-250BMoEQ8_0250B247.86 GiB6.75 GiB255.19 GiB128.81 GiB13±37%
Devstral-2-123B-Instruct-2512BF16125B232.89 GiB12.38 GiB245.97 GiB138.03 GiB3±8.3%
Mistral-Medium-3.5-128BBF16128B232.89 GiB12.38 GiB245.97 GiB138.03 GiB3±8.3%
Qwen3-VL-235B-A22B-InstructMoEQ8_0236B232.77 GiB6.61 GiB239.97 GiB144.03 GiB11±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ8_0236B232.77 GiB6.61 GiB239.97 GiB144.03 GiB11±37%
Qwen3-235B-A22BMoEQ8_0235B232.77 GiB6.61 GiB239.97 GiB144.03 GiB11±37%
Qwen3-235B-A22B-Thinking-2507MoEQ8_0235B232.77 GiB6.61 GiB239.97 GiB144.03 GiB11±37%
Qwen3-235B-A22B-Instruct-2507MoEQ8_0235B232.77 GiB6.61 GiB239.97 GiB144.03 GiB11±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ8_0236B233.41 GiB2.37 GiB236.37 GiB147.63 GiB15±37%
DeepSeek-V2.5MoEQ8_0236B233.41 GiB2.37 GiB236.37 GiB147.63 GiB15±37%
DeepSeek-Coder-V2-InstructMoEQ8_0236B233.41 GiB2.37 GiB236.37 GiB147.63 GiB15±37%
MiniMax-M2.1MoEQ8_0229B226.44 GiB8.72 GiB235.69 GiB148.31 GiB13±37%
MiniMax-M2.5MoEQ8_0229B226.44 GiB8.72 GiB235.69 GiB148.31 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 131,072 context with q4_0 KV cache, the largest being Step-3.7-Flash at BF16. 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.