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 32K 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 32K 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 GiB1.46 GiB191.98 GiB0.02 GiB19±37%
GLM-5MoEUD-IQ1_S754B189.71 GiB1.46 GiB191.76 GiB0.24 GiB19±37%
Hermes-4-405BQ3_K_M406B181.96 GiB8.37 GiB191.15 GiB0.85 GiB4±8.3%
Hermes-3-Llama-3.1-405BQ3_K_M406B181.96 GiB8.37 GiB191.15 GiB0.85 GiB4±8.3%
openPangu-2.0-FlashMoEKV unresolvedBF16100B188.04 GiB0.86 GiB189.45 GiB2.55 GiB20±37%
cogito-671b-v2.1MoEUD-IQ1_M671B187.19 GiB1.14 GiB188.96 GiB3.04 GiB20±37%
DeepSeek-V3.1-TerminusMoEUD-IQ1_M685B187.19 GiB1.14 GiB188.96 GiB3.04 GiB20±37%
DeepSeek-V3-0324MoEUD-IQ1_M685B186.94 GiB1.14 GiB188.71 GiB3.29 GiB20±37%
cogito-v2-preview-deepseek-671B-MoEMoEUD-IQ1_M671B186.85 GiB1.14 GiB188.62 GiB3.38 GiB20±37%
DeepSeek-R1-0528MoEUD-IQ1_M685B186.69 GiB1.14 GiB188.46 GiB3.54 GiB20±37%
DeepSeek-TNG-R1T2-ChimeraMoEUD-IQ1_M685B186.69 GiB1.14 GiB188.46 GiB3.54 GiB20±37%
Qwen3-Coder-480B-A35B-InstructMoEIQ3_XS480B183.57 GiB4.12 GiB188.27 GiB3.73 GiB15±37%
Qwen3-Coder-REAP-363B-A35BMoEIQ4_XS363B183.56 GiB4.12 GiB188.26 GiB3.74 GiB14±37%
MiniMax-M2.7MoEQ6_K229B183.52 GiB4.12 GiB188.17 GiB3.83 GiB18±37%
DeepSeek-Prover-V2-671BMoEUD-IQ1_M685B186.06 GiB1.14 GiB187.83 GiB4.17 GiB20±37%
DeepSeek-V3.2MoEUD-IQ1_M685B185.62 GiB1.14 GiB187.39 GiB4.61 GiB20±37%
GLM-4.5MoEIQ4_XS358B179.60 GiB6.11 GiB186.30 GiB5.70 GiB15±37%
Qwen3.5-397B-A17BMoEUD-IQ4_NL403B185.08 GiB0.50 GiB186.18 GiB5.82 GiB23±37%
GLM-4.7MoEIQ4_XS358B179.45 GiB6.11 GiB186.15 GiB5.85 GiB15±37%
MiniMax-M3MoEQ3_K_M427B183.11 GiB1.99 GiB185.67 GiB6.33 GiB19±37%
GLM-4.6-Derestricted-v3MoEIQ4_XS357B178.83 GiB6.11 GiB185.53 GiB6.47 GiB15±37%
GLM-4.6MoEIQ4_XS357B178.83 GiB6.11 GiB185.53 GiB6.47 GiB15±37%
r1-1776MoEIQ2_S671B183.47 GiB1.14 GiB185.24 GiB6.76 GiB20±37%
DeepSeek-R1MoEIQ2_S685B183.47 GiB1.14 GiB185.24 GiB6.76 GiB20±37%
GLM-4.6-REAP-268B-A32BMoEQ5_K_M269B177.71 GiB6.11 GiB184.41 GiB7.59 GiB13±37%
Ornith-1.0-397BMoEUD-IQ4_NL397B182.60 GiB0.50 GiB183.70 GiB8.30 GiB23±37%
MiMo-V2.5MoEKV unresolvedQ4_1311B180.99 GiB1.99 GiB183.58 GiB8.42 GiB20±37%
Qwen3-235B-A22BMoEQ6_K235B179.80 GiB3.12 GiB183.51 GiB8.49 GiB15±37%
Qwen3-235B-A22B-Instruct-2507MoEQ6_K235B179.80 GiB3.12 GiB183.51 GiB8.49 GiB15±37%
Qwen3-235B-A22B-Thinking-2507MoEQ6_K235B179.80 GiB3.12 GiB183.51 GiB8.49 GiB15±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ6_K236B179.76 GiB3.12 GiB183.46 GiB8.54 GiB15±37%
Qwen3-VL-235B-A22B-InstructMoEQ6_K236B179.76 GiB3.12 GiB183.46 GiB8.54 GiB15±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q6_K235B179.76 GiB3.12 GiB183.46 GiB8.54 GiB15±37%
grok-2MoEQ5_K_M270B178.41 GiB4.25 GiB183.36 GiB8.64 GiB7±37%
MiMo-V2-FlashMoEKV unresolvedQ4_1310B180.12 GiB1.99 GiB182.71 GiB9.29 GiB20±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ6_K236B180.25 GiB1.12 GiB181.96 GiB10.04 GiB19±37%
DeepSeek-V2.5MoEQ6_K236B180.25 GiB1.12 GiB181.96 GiB10.04 GiB19±37%
DeepSeek-Coder-V2-InstructMoEQ6_K236B180.25 GiB1.12 GiB181.96 GiB10.04 GiB19±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ3_K_M402B177.95 GiB3.19 GiB181.71 GiB10.29 GiB23±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 GiB1.14 GiB180.87 GiB11.13 GiB21±37%
Hy3MoEQ4_1299B174.90 GiB5.31 GiB180.80 GiB11.20 GiB16±37%
MiniMax-M2.1MoEQ6_K229B174.91 GiB4.12 GiB179.56 GiB12.44 GiB18±37%
MiniMax-M2MoEQ6_K229B174.91 GiB4.12 GiB179.56 GiB12.44 GiB18±37%
MiniMax-M2.5MoEQ6_K229B174.87 GiB4.12 GiB179.52 GiB12.48 GiB18±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ6_K229B174.87 GiB4.12 GiB179.52 GiB12.48 GiB18±37%
ERNIE-4.5-300B-A47B-PTQ4_1300B174.57 GiB3.59 GiB178.83 GiB13.17 GiB4±8.3%
Trinity-Large-PreviewMoEQ3_K_M399B176.43 GiB1.42 GiB178.42 GiB13.58 GiB24±37%
Trinity-Large-TrueBaseMoEI1-Q3_K_M399B176.30 GiB1.42 GiB178.30 GiB13.70 GiB24±37%
Trinity-Large-ThinkingMoEIQ3_M399B176.24 GiB1.42 GiB178.23 GiB13.77 GiB24±37%
command-a-plus-05-2026-bf16MoEQ6_K219B176.79 GiB0.76 GiB178.10 GiB13.90 GiB17±37%
Nex-N2-ProMoEIQ3_M397B176.93 GiB0.50 GiB178.03 GiB13.97 GiB24±37%
GLM-4.7-REAP-218B-A32BMoEQ6_K218B167.57 GiB6.11 GiB174.27 GiB17.73 GiB13±37%
Qwen3.5-REAP-262B-A17BMoEQ5_K_M262B172.96 GiB0.50 GiB174.06 GiB17.94 GiB22±37%
Minimax-M3-abliterated-cleanMoEQ3_K_S427B171.06 GiB1.99 GiB173.62 GiB18.38 GiB20±37%
GLM-5.2MoEQ3_K_M753B169.33 GiB1.46 GiB171.38 GiB20.62 GiB21±37%
Solar-Open2-250BMoEQ5_K_M250B165.65 GiB3.19 GiB169.41 GiB22.59 GiB21±37%
Step-3.7-FlashQ6_K_L201B160.20 GiB6.92 GiB167.70 GiB24.30 GiB4±8.3%
dots.llm1.instMoEQ8_0143B141.37 GiB16.47 GiB158.42 GiB33.58 GiB12±37%
step-3.5-flashQ6_K199B150.81 GiB6.92 GiB158.31 GiB33.69 GiB4±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 32,768 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.