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 64K 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 64K 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 GiB1.21 GiB382.96 GiB1.04 GiB11±37%
DeepSeek-V3.1-TerminusMoEQ4_K_M685B381.12 GiB1.21 GiB382.96 GiB1.04 GiB11±37%
DeepSeek-V3.1MoEQ4_K_M685B381.12 GiB1.21 GiB382.96 GiB1.04 GiB11±37%
DeepSeek-V3.2MoEQ4_K_M685B377.56 GiB1.21 GiB379.39 GiB4.61 GiB11±37%
cogito-671b-v2.1MoEQ4_K_M671B377.55 GiB1.21 GiB379.39 GiB4.61 GiB11±37%
Kimi-K2.5MoEIQ3_S1059B377.51 GiB1.21 GiB379.35 GiB4.65 GiB12±37%
DeepSeek-TNG-R1T2-ChimeraMoEQ4_K_M685B377.13 GiB1.21 GiB378.97 GiB5.03 GiB11±37%
cogito-v2-preview-deepseek-671B-MoEMoEQ4_K_M671B377.13 GiB1.21 GiB378.96 GiB5.04 GiB11±37%
DeepSeek-V3-0324MoEQ4_K_M685B376.89 GiB1.21 GiB378.73 GiB5.27 GiB11±37%
DeepSeek-Prover-V2-671BMoEQ4_K_M685B376.71 GiB1.21 GiB378.55 GiB5.45 GiB11±37%
r1-1776MoEQ4_K_M671B376.65 GiB1.21 GiB378.49 GiB5.51 GiB11±37%
DeepSeek-R1MoEQ4_K_M685B376.65 GiB1.21 GiB378.49 GiB5.51 GiB11±37%
GLM-5.1MoEIQ4_XS754B375.69 GiB1.54 GiB377.83 GiB6.17 GiB11±37%
GLM-5MoEIQ4_XS754B375.23 GiB1.54 GiB377.37 GiB6.63 GiB11±37%
Step-3.7-FlashBF16201B366.96 GiB7.04 GiB374.58 GiB9.42 GiB2±8.3%
Qwen3-Coder-480B-A35B-InstructMoEQ6_K480B367.10 GiB4.36 GiB372.04 GiB11.96 GiB9±37%
Kimi-K2-ThinkingMoEIQ3_XXS1058B367.09 GiB1.21 GiB368.92 GiB15.08 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 GiB4.36 GiB364.39 GiB19.61 GiB8±37%
GLM-4.5MoEQ8_0358B354.80 GiB6.47 GiB361.86 GiB22.14 GiB9±37%
GLM-4.7MoEQ8_0358B354.80 GiB6.47 GiB361.86 GiB22.14 GiB9±37%
GLM-4.6MoEQ8_0357B353.27 GiB6.47 GiB360.33 GiB23.67 GiB9±37%
GLM-4.6-Derestricted-v3MoEQ8_0357B353.27 GiB6.47 GiB360.33 GiB23.67 GiB9±37%
MiMo-V2.5-ProMoEKV unresolvedUD-IQ3_S1023B351.94 GiB6.15 GiB358.71 GiB25.29 GiB11±37%
Kimi-K2.7-CodeMoEUD-IQ3_XXS1059B351.00 GiB1.21 GiB352.83 GiB31.17 GiB13±37%
Kimi-K2-InstructMoEQ2_K_L1026B347.81 GiB1.21 GiB349.64 GiB34.36 GiB13±37%
GLM-5.2MoEUD-IQ4_NL753B347.07 GiB1.54 GiB349.21 GiB34.79 GiB12±37%
MiniMax-M3MoEQ6_K427B344.03 GiB2.11 GiB346.71 GiB37.29 GiB11±37%
Kimi-K2.6MoEQ2_K_L1059B334.66 GiB1.21 GiB336.49 GiB47.51 GiB13±37%
Qwen3.5-397B-A17BMoEQ6_K_L403B325.38 GiB0.53 GiB326.50 GiB57.50 GiB14±37%
Trinity-Large-ThinkingMoEQ6_K_L399B319.94 GiB1.28 GiB321.79 GiB62.21 GiB15±37%
Ornith-1.0-397BMoEQ6_K397B318.87 GiB0.53 GiB320.00 GiB64.00 GiB14±37%
Nex-N2-ProMoEQ6_K397B318.87 GiB0.53 GiB320.00 GiB64.00 GiB14±37%
Hermes-3-Llama-3.1-405BQ6_K406B310.08 GiB8.86 GiB319.77 GiB64.23 GiB2±8.3%
Hermes-4-405BQ6_K406B310.08 GiB8.86 GiB319.77 GiB64.23 GiB2±8.3%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ6_K402B306.19 GiB3.38 GiB310.15 GiB73.85 GiB15±37%
MiMo-V2.5MoEKV unresolvedQ8_0311B306.67 GiB2.11 GiB309.37 GiB74.63 GiB13±37%
MiMo-V2-FlashMoEKV unresolvedQ8_0310B305.69 GiB2.11 GiB308.40 GiB75.60 GiB13±37%
Trinity-Large-PreviewMoEQ6_K_L399B305.35 GiB1.28 GiB307.20 GiB76.80 GiB15±37%
Trinity-Large-TrueBaseMoEQ6_K399B305.07 GiB1.28 GiB306.92 GiB77.08 GiB15±37%
Hy3MoEQ8_0299B295.84 GiB5.63 GiB302.05 GiB81.95 GiB11±37%
ERNIE-4.5-300B-A47B-PTQ8_0300B296.43 GiB3.80 GiB300.90 GiB83.10 GiB2±8.3%
dots.llm1.instMoEBF16143B266.00 GiB17.44 GiB284.02 GiB99.98 GiB8±37%
GLM-4.6-REAP-268B-A32BMoEQ8_0269B266.14 GiB6.47 GiB273.20 GiB110.80 GiB10±37%
grok-2MoEQ8_0270B266.72 GiB4.50 GiB271.91 GiB112.09 GiB5±37%
InklingMoEUD-IQ1_M952B265.46 GiB4.64 GiB270.67 GiB113.33 GiB14±37%
Athene-70BF3270.6B262.84 GiB5.63 GiB269.14 GiB114.86 GiB3±8.3%
Solar-Open2-250BMoEQ8_0250B247.86 GiB3.38 GiB251.81 GiB132.19 GiB15±37%
Devstral-2-123B-Instruct-2512BF16125B232.89 GiB6.19 GiB239.79 GiB144.21 GiB3±8.3%
Mistral-Medium-3.5-128BBF16128B232.89 GiB6.19 GiB239.79 GiB144.21 GiB3±8.3%
Qwen3-VL-235B-A22B-InstructMoEQ8_0236B232.77 GiB3.30 GiB236.66 GiB147.34 GiB12±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ8_0236B232.77 GiB3.30 GiB236.66 GiB147.34 GiB12±37%
Qwen3-235B-A22BMoEQ8_0235B232.77 GiB3.30 GiB236.66 GiB147.34 GiB12±37%
Qwen3-235B-A22B-Thinking-2507MoEQ8_0235B232.77 GiB3.30 GiB236.66 GiB147.34 GiB12±37%
Qwen3-235B-A22B-Instruct-2507MoEQ8_0235B232.77 GiB3.30 GiB236.66 GiB147.34 GiB12±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ8_0236B233.41 GiB1.19 GiB235.19 GiB148.81 GiB16±37%
DeepSeek-V2.5MoEQ8_0236B233.41 GiB1.19 GiB235.19 GiB148.81 GiB16±37%
DeepSeek-Coder-V2-InstructMoEQ8_0236B233.41 GiB1.19 GiB235.19 GiB148.81 GiB16±37%
Qwen3.5-122B-A10B-hereticMoEBF16123B232.24 GiB0.42 GiB233.24 GiB150.76 GiB17±37%
Qwen3.5-122B-A10BMoEBF16125B232.24 GiB0.42 GiB233.24 GiB150.76 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 65,536 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.