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 4K 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 4K context

largest quantization that fits, per model · 2108 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Qwen3-Coder-REAP-363B-A35BMoEIQ4_NL363B190.65 GiB0.51 GiB191.75 GiB0.25 GiB15±37%
MiniMax-M3MoEQ3_K_L427B190.89 GiB0.25 GiB191.71 GiB0.29 GiB20±37%
GLM-4.5MoEIQ4_NL358B189.74 GiB0.76 GiB191.10 GiB0.90 GiB17±37%
GLM-4.7MoEIQ4_NL358B189.62 GiB0.76 GiB190.97 GiB1.03 GiB17±37%
GLM-5.1MoEIQ2_XXS754B189.92 GiB0.18 GiB190.70 GiB1.30 GiB21±37%
GLM-5MoEUD-IQ1_S754B189.71 GiB0.18 GiB190.49 GiB1.51 GiB21±37%
GLM-4.6-Derestricted-v3MoEIQ4_NL357B188.93 GiB0.76 GiB190.28 GiB1.72 GiB17±37%
GLM-4.6MoEIQ4_NL357B188.93 GiB0.76 GiB190.28 GiB1.72 GiB17±37%
Qwen3-Coder-480B-A35B-InstructMoEUD-IQ3_XXS480B187.70 GiB0.51 GiB188.80 GiB3.20 GiB17±37%
openPangu-2.0-FlashMoEKV unresolvedBF16100B188.04 GiB0.11 GiB188.70 GiB3.30 GiB21±37%
cogito-671b-v2.1MoEUD-IQ1_M671B187.19 GiB0.14 GiB187.96 GiB4.04 GiB21±37%
DeepSeek-V3.1-TerminusMoEUD-IQ1_M685B187.19 GiB0.14 GiB187.96 GiB4.04 GiB21±37%
DeepSeek-V3-0324MoEUD-IQ1_M685B186.94 GiB0.14 GiB187.71 GiB4.29 GiB21±37%
cogito-v2-preview-deepseek-671B-MoEMoEUD-IQ1_M671B186.85 GiB0.14 GiB187.62 GiB4.38 GiB21±37%
DeepSeek-R1-0528MoEUD-IQ1_M685B186.69 GiB0.14 GiB187.46 GiB4.54 GiB21±37%
DeepSeek-TNG-R1T2-ChimeraMoEUD-IQ1_M685B186.69 GiB0.14 GiB187.46 GiB4.54 GiB21±37%
DeepSeek-Prover-V2-671BMoEUD-IQ1_M685B186.06 GiB0.14 GiB186.83 GiB5.17 GiB21±37%
DeepSeek-V3.2MoEUD-IQ1_M685B185.62 GiB0.14 GiB186.39 GiB5.61 GiB21±37%
Qwen3.5-397B-A17BMoEUD-IQ4_NL403B185.08 GiB0.06 GiB185.74 GiB6.26 GiB24±37%
MiniMax-M2.7MoEQ6_K229B183.52 GiB0.51 GiB184.57 GiB7.43 GiB21±37%
r1-1776MoEIQ2_S671B183.47 GiB0.14 GiB184.24 GiB7.76 GiB21±37%
DeepSeek-R1MoEIQ2_S685B183.47 GiB0.14 GiB184.24 GiB7.76 GiB21±37%
Hermes-4-405BQ3_K_M406B181.96 GiB1.05 GiB183.83 GiB8.17 GiB4±8.3%
Hermes-3-Llama-3.1-405BQ3_K_M406B181.96 GiB1.05 GiB183.83 GiB8.17 GiB4±8.3%
Ornith-1.0-397BMoEUD-IQ4_NL397B182.60 GiB0.06 GiB183.26 GiB8.74 GiB24±37%
MiMo-V2.5MoEKV unresolvedQ4_1311B180.99 GiB0.25 GiB181.83 GiB10.17 GiB21±37%
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16MoEUD-IQ2_M561B180.64 GiB0.00 GiB181.23 GiB10.77 GiB20±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ6_K236B180.25 GiB0.14 GiB180.98 GiB11.02 GiB20±37%
DeepSeek-V2.5MoEQ6_K236B180.25 GiB0.14 GiB180.98 GiB11.02 GiB20±37%
DeepSeek-Coder-V2-InstructMoEQ6_K236B180.25 GiB0.14 GiB180.98 GiB11.02 GiB20±37%
MiMo-V2-FlashMoEKV unresolvedQ4_1310B180.12 GiB0.25 GiB180.97 GiB11.03 GiB21±37%
Qwen3-235B-A22BMoEQ6_K235B179.80 GiB0.39 GiB180.78 GiB11.22 GiB17±37%
Qwen3-235B-A22B-Instruct-2507MoEQ6_K235B179.80 GiB0.39 GiB180.78 GiB11.22 GiB17±37%
Qwen3-235B-A22B-Thinking-2507MoEQ6_K235B179.80 GiB0.39 GiB180.78 GiB11.22 GiB17±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ6_K236B179.76 GiB0.39 GiB180.73 GiB11.27 GiB17±37%
Qwen3-VL-235B-A22B-InstructMoEQ6_K236B179.76 GiB0.39 GiB180.73 GiB11.27 GiB17±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q6_K235B179.76 GiB0.39 GiB180.73 GiB11.27 GiB17±37%
DeepSeek-V3.1MoEUD-IQ1_S685B179.11 GiB0.14 GiB179.88 GiB12.12 GiB22±37%
grok-2MoEQ5_K_M270B178.41 GiB0.53 GiB179.64 GiB12.36 GiB7±37%
GLM-4.6-REAP-268B-A32BMoEQ5_K_M269B177.71 GiB0.76 GiB179.07 GiB12.93 GiB16±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ3_K_M402B177.95 GiB0.40 GiB178.92 GiB13.08 GiB27±37%
command-a-plus-05-2026-bf16MoEQ6_K219B176.79 GiB0.27 GiB177.61 GiB14.39 GiB17±37%
Nex-N2-ProMoEIQ3_M397B176.93 GiB0.06 GiB177.60 GiB14.40 GiB25±37%
Trinity-Large-PreviewMoEQ3_K_M399B176.43 GiB0.50 GiB177.50 GiB14.50 GiB25±37%
Trinity-Large-TrueBaseMoEI1-Q3_K_M399B176.30 GiB0.50 GiB177.38 GiB14.62 GiB25±37%
Trinity-Large-ThinkingMoEIQ3_M399B176.24 GiB0.50 GiB177.32 GiB14.68 GiB25±37%
Hy3MoEQ4_1299B174.90 GiB0.66 GiB176.15 GiB15.85 GiB20±37%
MiniMax-M2.1MoEQ6_K229B174.91 GiB0.51 GiB175.96 GiB16.04 GiB22±37%
MiniMax-M2MoEQ6_K229B174.91 GiB0.51 GiB175.96 GiB16.04 GiB22±37%
MiniMax-M2.5MoEQ6_K229B174.87 GiB0.51 GiB175.92 GiB16.08 GiB22±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ6_K229B174.87 GiB0.51 GiB175.92 GiB16.08 GiB22±37%
ERNIE-4.5-300B-A47B-PTQ4_1300B174.57 GiB0.45 GiB175.69 GiB16.31 GiB4±8.3%
Qwen3.5-REAP-262B-A17BMoEQ5_K_M262B172.96 GiB0.06 GiB173.62 GiB18.38 GiB23±37%
Minimax-M3-abliterated-cleanMoEQ3_K_S427B171.06 GiB0.25 GiB171.88 GiB20.12 GiB22±37%
GLM-5.2MoEQ3_K_M753B169.33 GiB0.18 GiB170.11 GiB21.89 GiB23±37%
GLM-4.7-REAP-218B-A32BMoEQ6_K218B167.57 GiB0.76 GiB168.92 GiB23.08 GiB15±37%
Solar-Open2-250BMoEQ5_K_M250B165.65 GiB0.40 GiB166.62 GiB25.38 GiB24±37%
Step-3.7-FlashQ6_K_L201B160.20 GiB1.34 GiB162.12 GiB29.88 GiB4±8.3%
DeepSeek-V4-FlashMoEQ4_K291B153.33 GiB0.03 GiB153.96 GiB38.04 GiB27±37%
step-3.5-flashQ6_K199B150.81 GiB1.34 GiB152.74 GiB39.26 GiB5±8.3%
●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 4,096 context with q8_0 KV cache, the largest being Qwen3-Coder-REAP-363B-A35B at IQ4_NL. 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.