Apple · apple

Apple M3 Ultra

Apple M3 Ultra has 256 GB of unified memory at 819 GB/s — about 178.56 GiB usable after driver and compositor overhead. 2108 of 2118 indexed models fit at 128K context with q8_0 KV. Note only 192 GB of its 256 GB is allocatable to the GPU.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
256 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 1812vision language 192image 2audio tts 21audio asr 39video 16embedding 26

What fits at 128K context

largest quantization that fits, per model · 2108 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
MiniMax-M2.1MoEQ6_K229B174.91 GiB16.47 GiB191.92 GiB0.08 GiB12±37%
MiniMax-M2MoEQ6_K229B174.91 GiB16.47 GiB191.92 GiB0.08 GiB12±37%
DeepSeek-R1-0528MoEUD-IQ1_M685B186.69 GiB4.56 GiB191.87 GiB0.13 GiB17±37%
DeepSeek-TNG-R1T2-ChimeraMoEUD-IQ1_M685B186.69 GiB4.56 GiB191.87 GiB0.13 GiB17±37%
MiniMax-M2.5MoEQ6_K229B174.87 GiB16.47 GiB191.87 GiB0.13 GiB12±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ6_K229B174.87 GiB16.47 GiB191.87 GiB0.13 GiB12±37%
Hy3MoEQ4_K_L299B169.99 GiB21.25 GiB191.83 GiB0.17 GiB10±37%
Llama-3_3-Nemotron-Super-49B-v1_5IQ3_M49.9B21.10 GiB170.00 GiB191.79 GiB0.21 GiB4±8.3%
Valkyrie-49B-v2.1I1-IQ3_M49.9B21.10 GiB170.00 GiB191.79 GiB0.21 GiB4±8.3%
Llama-3_3-Nemotron-Super-49B-v1IQ3_M49.9B21.10 GiB170.00 GiB191.79 GiB0.21 GiB4±8.3%
Llama-3_1-Nemotron-51B-InstructQ3_K_S51.5B21.10 GiB170.00 GiB191.79 GiB0.21 GiB4±8.3%
MiniMax-M3MoEQ3_K_M427B183.11 GiB7.97 GiB191.64 GiB0.36 GiB15±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ3_K_M402B177.95 GiB12.75 GiB191.27 GiB0.73 GiB15±37%
DeepSeek-Prover-V2-671BMoEUD-IQ1_M685B186.06 GiB4.56 GiB191.25 GiB0.75 GiB17±37%
DeepSeek-V3.2MoEUD-IQ1_M685B185.62 GiB4.56 GiB190.81 GiB1.19 GiB17±37%
grok-2MoEQ5_K_S270B173.10 GiB17.00 GiB190.79 GiB1.21 GiB6±37%
ERNIE-4.5-300B-A47B-PTQ4_1300B174.57 GiB14.34 GiB189.58 GiB2.42 GiB4±8.3%
MiMo-V2.5MoEKV unresolvedQ4_1311B180.99 GiB7.97 GiB189.55 GiB2.45 GiB15±37%
MiMo-V2-FlashMoEKV unresolvedQ4_1310B180.12 GiB7.97 GiB188.69 GiB3.31 GiB15±37%
DeepSeek-V3-0324MoEIQ2_S685B183.47 GiB4.56 GiB188.65 GiB3.35 GiB17±37%
r1-1776MoEIQ2_S671B183.47 GiB4.56 GiB188.65 GiB3.35 GiB17±37%
DeepSeek-R1MoEIQ2_S685B183.47 GiB4.56 GiB188.65 GiB3.35 GiB17±37%
Qwen3.5-397B-A17BMoEUD-IQ4_NL403B185.08 GiB1.99 GiB187.67 GiB4.33 GiB21±37%
DeepSeek-V3.1-TerminusMoEIQ2_XS685B181.82 GiB4.56 GiB187.01 GiB4.99 GiB17±37%
Step-3.7-FlashQ6_K_L201B160.20 GiB26.05 GiB186.83 GiB5.17 GiB4±8.3%
dots.llm1.instMoEQ6_K143B120.06 GiB65.88 GiB186.51 GiB5.49 GiB5±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ6_K236B180.25 GiB4.48 GiB185.32 GiB6.68 GiB17±37%
DeepSeek-V2.5MoEQ6_K236B180.25 GiB4.48 GiB185.32 GiB6.68 GiB17±37%
DeepSeek-Coder-V2-InstructMoEQ6_K236B180.25 GiB4.48 GiB185.32 GiB6.68 GiB17±37%
Ornith-1.0-397BMoEUD-IQ4_NL397B182.60 GiB1.99 GiB185.19 GiB6.81 GiB21±37%
GLM-4.5MoEQ3_K_M358B159.51 GiB24.44 GiB184.53 GiB7.47 GiB9±37%
GLM-4.7MoEQ3_K_M358B159.51 GiB24.44 GiB184.53 GiB7.47 GiB9±37%
DeepSeek-V3.1MoEUD-IQ1_S685B179.11 GiB4.56 GiB184.29 GiB7.71 GiB18±37%
GLM-4.6MoEQ3_K_M357B158.89 GiB24.44 GiB183.91 GiB8.09 GiB9±37%
GLM-4.6-Derestricted-v3MoEQ3_K_L357B158.25 GiB24.44 GiB183.28 GiB8.72 GiB9±37%
GLM-4.6-REAP-268B-A32BMoEQ4_1269B156.99 GiB24.44 GiB182.02 GiB9.98 GiB9±37%
Trinity-Large-PreviewMoEQ3_K_M399B176.43 GiB4.40 GiB181.41 GiB10.59 GiB20±37%
Trinity-Large-TrueBaseMoEI1-Q3_K_M399B176.30 GiB4.40 GiB181.29 GiB10.71 GiB20±37%
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16MoEUD-IQ2_M561B180.64 GiB0.00 GiB181.23 GiB10.77 GiB20±37%
Trinity-Large-ThinkingMoEIQ3_M399B176.24 GiB4.40 GiB181.22 GiB10.78 GiB20±37%
Qwen3-Coder-480B-A35B-InstructMoEQ2_K_L480B162.87 GiB16.47 GiB179.92 GiB12.08 GiB11±37%
command-a-plus-05-2026-bf16MoEQ6_K219B176.79 GiB2.35 GiB179.70 GiB12.30 GiB16±37%
Minimax-M3-abliterated-cleanMoEQ3_K_S427B171.06 GiB7.97 GiB179.60 GiB12.40 GiB16±37%
Nex-N2-ProMoEIQ3_M397B176.93 GiB1.99 GiB179.53 GiB12.47 GiB22±37%
Hermes-3-Llama-3.1-405BIQ3_XXS406B145.14 GiB33.47 GiB179.44 GiB12.56 GiB4±8.3%
Solar-Open2-250BMoEQ5_K_M250B165.65 GiB12.75 GiB178.97 GiB13.03 GiB14±37%
Qwen3-Coder-REAP-363B-A35BMoEQ3_K_M363B161.57 GiB16.47 GiB178.63 GiB13.37 GiB10±37%
cogito-v2-preview-deepseek-671B-MoEMoEUD-IQ1_S671B173.12 GiB4.56 GiB178.31 GiB13.69 GiB18±37%
cogito-671b-v2.1MoEUD-IQ1_S671B173.11 GiB4.56 GiB178.30 GiB13.70 GiB18±37%
step-3.5-flashQ6_K199B150.81 GiB26.05 GiB177.44 GiB14.56 GiB4±8.3%
GLM-5.2MoEQ3_K_M753B169.33 GiB5.83 GiB175.75 GiB16.25 GiB17±37%
Qwen3.5-REAP-262B-A17BMoEQ5_K_M262B172.96 GiB1.99 GiB175.55 GiB16.45 GiB20±37%
MiniMax-M2.7MoEUD-Q5_K_M229B157.23 GiB16.47 GiB174.24 GiB17.76 GiB12±37%
Hermes-4-405BQ2_K406B139.07 GiB33.47 GiB173.37 GiB18.63 GiB4±8.3%
HarmonicHarlequin_v5-20BQ8_033.3B32.97 GiB138.13 GiB171.69 GiB20.31 GiB4±8.3%
GLM-5MoEUD-TQ1_0754B164.05 GiB5.83 GiB170.47 GiB21.53 GiB18±37%
GLM-5.1MoEIQ1_M754B163.93 GiB5.83 GiB170.36 GiB21.64 GiB18±37%
GLM-4.7-REAP-218B-A32BMoEQ5_K_M218B145.23 GiB24.44 GiB170.26 GiB21.74 GiB9±37%
Qwen3-235B-A22B-Thinking-2507MoEQ5_K_M235B155.43 GiB12.48 GiB168.49 GiB23.51 GiB12±37%
Qwen3-235B-A22B-Instruct-2507MoEQ5_K_M235B155.43 GiB12.48 GiB168.49 GiB23.51 GiB12±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?
2108 of 2118 indexed open-weight models fit a Apple M3 Ultra at 131,072 context with q8_0 KV cache, the largest being MiniMax-M2.1 at Q6_K. That covers text, vision-language, image, video and speech models.
How much usable memory does a Apple M3 Ultra actually have?
Its nameplate is 256 GB, but about 178.56 GiB is available to a model once driver and compositor overhead is accounted for, and only 192 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.