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 16K 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 16K context

largest quantization that fits, per model · 2108 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
GLM-5.1MoEIQ2_XXS754B189.92 GiB0.73 GiB191.25 GiB0.75 GiB20±37%
GLM-5MoEUD-IQ1_S754B189.71 GiB0.73 GiB191.04 GiB0.96 GiB20±37%
Qwen3-Coder-480B-A35B-InstructMoEUD-IQ3_XXS480B187.70 GiB2.06 GiB190.34 GiB1.66 GiB16±37%
openPangu-2.0-FlashMoEKV unresolvedBF16100B188.04 GiB0.43 GiB189.03 GiB2.97 GiB20±37%
cogito-671b-v2.1MoEUD-IQ1_M671B187.19 GiB0.57 GiB188.39 GiB3.61 GiB20±37%
DeepSeek-V3.1-TerminusMoEUD-IQ1_M685B187.19 GiB0.57 GiB188.39 GiB3.61 GiB20±37%
DeepSeek-V3-0324MoEUD-IQ1_M685B186.94 GiB0.57 GiB188.14 GiB3.86 GiB20±37%
cogito-v2-preview-deepseek-671B-MoEMoEUD-IQ1_M671B186.85 GiB0.57 GiB188.05 GiB3.95 GiB20±37%
DeepSeek-R1-0528MoEUD-IQ1_M685B186.69 GiB0.57 GiB187.89 GiB4.11 GiB20±37%
DeepSeek-TNG-R1T2-ChimeraMoEUD-IQ1_M685B186.69 GiB0.57 GiB187.89 GiB4.11 GiB20±37%
DeepSeek-Prover-V2-671BMoEUD-IQ1_M685B186.06 GiB0.57 GiB187.26 GiB4.74 GiB21±37%
Hermes-4-405BQ3_K_M406B181.96 GiB4.18 GiB186.97 GiB5.03 GiB4±8.3%
Hermes-3-Llama-3.1-405BQ3_K_M406B181.96 GiB4.18 GiB186.97 GiB5.03 GiB4±8.3%
DeepSeek-V3.2MoEUD-IQ1_M685B185.62 GiB0.57 GiB186.82 GiB5.18 GiB21±37%
Qwen3-Coder-REAP-363B-A35BMoEIQ4_XS363B183.56 GiB2.06 GiB186.20 GiB5.80 GiB15±37%
MiniMax-M2.7MoEQ6_K229B183.52 GiB2.06 GiB186.12 GiB5.88 GiB19±37%
Qwen3.5-397B-A17BMoEUD-IQ4_NL403B185.08 GiB0.25 GiB185.93 GiB6.07 GiB23±37%
MiniMax-M3MoEQ3_K_M427B183.11 GiB1.00 GiB184.67 GiB7.33 GiB20±37%
r1-1776MoEIQ2_S671B183.47 GiB0.57 GiB184.67 GiB7.33 GiB21±37%
DeepSeek-R1MoEIQ2_S685B183.47 GiB0.57 GiB184.67 GiB7.33 GiB21±37%
Ornith-1.0-397BMoEUD-IQ4_NL397B182.60 GiB0.25 GiB183.45 GiB8.55 GiB24±37%
GLM-4.5MoEIQ4_XS358B179.60 GiB3.05 GiB183.25 GiB8.75 GiB16±37%
GLM-4.7MoEIQ4_XS358B179.45 GiB3.05 GiB183.09 GiB8.91 GiB16±37%
MiMo-V2.5MoEKV unresolvedQ4_1311B180.99 GiB1.00 GiB182.58 GiB9.42 GiB21±37%
GLM-4.6-Derestricted-v3MoEIQ4_XS357B178.83 GiB3.05 GiB182.48 GiB9.52 GiB16±37%
GLM-4.6MoEIQ4_XS357B178.83 GiB3.05 GiB182.48 GiB9.52 GiB16±37%
Qwen3-235B-A22BMoEQ6_K235B179.80 GiB1.56 GiB181.95 GiB10.05 GiB16±37%
Qwen3-235B-A22B-Instruct-2507MoEQ6_K235B179.80 GiB1.56 GiB181.95 GiB10.05 GiB16±37%
Qwen3-235B-A22B-Thinking-2507MoEQ6_K235B179.80 GiB1.56 GiB181.95 GiB10.05 GiB16±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ6_K236B179.76 GiB1.56 GiB181.90 GiB10.10 GiB16±37%
Qwen3-VL-235B-A22B-InstructMoEQ6_K236B179.76 GiB1.56 GiB181.90 GiB10.10 GiB16±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q6_K235B179.76 GiB1.56 GiB181.90 GiB10.10 GiB16±37%
MiMo-V2-FlashMoEKV unresolvedQ4_1310B180.12 GiB1.00 GiB181.71 GiB10.29 GiB21±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ6_K236B180.25 GiB0.56 GiB181.40 GiB10.60 GiB20±37%
DeepSeek-V2.5MoEQ6_K236B180.25 GiB0.56 GiB181.40 GiB10.60 GiB20±37%
DeepSeek-Coder-V2-InstructMoEQ6_K236B180.25 GiB0.56 GiB181.40 GiB10.60 GiB20±37%
GLM-4.6-REAP-268B-A32BMoEQ5_K_M269B177.71 GiB3.05 GiB181.36 GiB10.64 GiB15±37%
grok-2MoEQ5_K_M270B178.41 GiB2.13 GiB181.23 GiB10.77 GiB7±37%
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16MoEUD-IQ2_M561B180.64 GiB0.00 GiB181.23 GiB10.77 GiB20±37%
DeepSeek-V3.1MoEUD-IQ1_S685B179.11 GiB0.57 GiB180.31 GiB11.69 GiB21±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ3_K_M402B177.95 GiB1.59 GiB180.12 GiB11.88 GiB25±37%
Hy3MoEQ4_1299B174.90 GiB2.66 GiB178.14 GiB13.86 GiB18±37%
Trinity-Large-PreviewMoEQ3_K_M399B176.43 GiB0.92 GiB177.92 GiB14.08 GiB24±37%
command-a-plus-05-2026-bf16MoEQ6_K219B176.79 GiB0.49 GiB177.84 GiB14.16 GiB17±37%
Trinity-Large-TrueBaseMoEI1-Q3_K_M399B176.30 GiB0.92 GiB177.80 GiB14.20 GiB24±37%
Nex-N2-ProMoEIQ3_M397B176.93 GiB0.25 GiB177.78 GiB14.22 GiB24±37%
Trinity-Large-ThinkingMoEIQ3_M399B176.24 GiB0.92 GiB177.74 GiB14.26 GiB24±37%
MiniMax-M2.1MoEQ6_K229B174.91 GiB2.06 GiB177.51 GiB14.49 GiB20±37%
MiniMax-M2MoEQ6_K229B174.91 GiB2.06 GiB177.51 GiB14.49 GiB20±37%
MiniMax-M2.5MoEQ6_K229B174.87 GiB2.06 GiB177.46 GiB14.54 GiB20±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ6_K229B174.87 GiB2.06 GiB177.46 GiB14.54 GiB20±37%
ERNIE-4.5-300B-A47B-PTQ4_1300B174.57 GiB1.79 GiB177.03 GiB14.97 GiB4±8.3%
Qwen3.5-REAP-262B-A17BMoEQ5_K_M262B172.96 GiB0.25 GiB173.81 GiB18.19 GiB22±37%
Minimax-M3-abliterated-cleanMoEQ3_K_S427B171.06 GiB1.00 GiB172.62 GiB19.38 GiB21±37%
GLM-4.7-REAP-218B-A32BMoEQ6_K218B167.57 GiB3.05 GiB171.21 GiB20.79 GiB14±37%
GLM-5.2MoEQ3_K_M753B169.33 GiB0.73 GiB170.65 GiB21.35 GiB22±37%
Solar-Open2-250BMoEQ5_K_M250B165.65 GiB1.59 GiB167.82 GiB24.18 GiB23±37%
Step-3.7-FlashQ6_K_L201B160.20 GiB3.74 GiB164.51 GiB27.49 GiB4±8.3%
step-3.5-flashQ6_K199B150.81 GiB3.74 GiB155.13 GiB36.87 GiB4±8.3%
DeepSeek-V4-FlashMoEQ4_K291B153.33 GiB0.03 GiB153.96 GiB38.04 GiB27±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 16,384 context with q8_0 KV cache, the largest being GLM-5.1 at IQ2_XXS. 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.