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 64K context with q4_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 64K context

largest quantization that fits, per model · 2108 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
GLM-5MoEUD-IQ1_S754B189.71 GiB1.54 GiB191.85 GiB0.15 GiB19±37%
Hermes-4-405BQ3_K_M406B181.96 GiB8.86 GiB191.65 GiB0.35 GiB4±8.3%
Hermes-3-Llama-3.1-405BQ3_K_M406B181.96 GiB8.86 GiB191.65 GiB0.35 GiB4±8.3%
openPangu-2.0-FlashMoEKV unresolvedBF16100B188.04 GiB0.91 GiB189.51 GiB2.49 GiB20±37%
cogito-671b-v2.1MoEUD-IQ1_M671B187.19 GiB1.21 GiB189.03 GiB2.97 GiB20±37%
DeepSeek-V3.1-TerminusMoEUD-IQ1_M685B187.19 GiB1.21 GiB189.02 GiB2.98 GiB20±37%
DeepSeek-V3-0324MoEUD-IQ1_M685B186.94 GiB1.21 GiB188.78 GiB3.22 GiB20±37%
cogito-v2-preview-deepseek-671B-MoEMoEUD-IQ1_M671B186.85 GiB1.21 GiB188.68 GiB3.32 GiB20±37%
DeepSeek-R1-0528MoEUD-IQ1_M685B186.69 GiB1.21 GiB188.52 GiB3.48 GiB20±37%
DeepSeek-TNG-R1T2-ChimeraMoEUD-IQ1_M685B186.69 GiB1.21 GiB188.52 GiB3.48 GiB20±37%
Qwen3-Coder-480B-A35B-InstructMoEIQ3_XS480B183.57 GiB4.36 GiB188.51 GiB3.49 GiB15±37%
Qwen3-Coder-REAP-363B-A35BMoEIQ4_XS363B183.56 GiB4.36 GiB188.50 GiB3.50 GiB14±37%
MiniMax-M2.7MoEQ6_K229B183.52 GiB4.36 GiB188.42 GiB3.58 GiB17±37%
DeepSeek-Prover-V2-671BMoEUD-IQ1_M685B186.06 GiB1.21 GiB187.90 GiB4.10 GiB20±37%
DeepSeek-V3.2MoEUD-IQ1_M685B185.62 GiB1.21 GiB187.46 GiB4.54 GiB20±37%
GLM-4.5MoEIQ4_XS358B179.60 GiB6.47 GiB186.66 GiB5.34 GiB14±37%
GLM-4.7MoEIQ4_XS358B179.45 GiB6.47 GiB186.50 GiB5.50 GiB14±37%
Qwen3.5-397B-A17BMoEUD-IQ4_NL403B185.08 GiB0.53 GiB186.21 GiB5.79 GiB23±37%
GLM-4.6-Derestricted-v3MoEIQ4_XS357B178.83 GiB6.47 GiB185.89 GiB6.11 GiB14±37%
GLM-4.6MoEIQ4_XS357B178.83 GiB6.47 GiB185.89 GiB6.11 GiB14±37%
MiniMax-M3MoEQ3_K_M427B183.11 GiB2.11 GiB185.79 GiB6.21 GiB19±37%
r1-1776MoEIQ2_S671B183.47 GiB1.21 GiB185.30 GiB6.70 GiB20±37%
DeepSeek-R1MoEIQ2_S685B183.47 GiB1.21 GiB185.30 GiB6.70 GiB20±37%
GLM-4.6-REAP-268B-A32BMoEQ5_K_M269B177.71 GiB6.47 GiB184.77 GiB7.23 GiB13±37%
Ornith-1.0-397BMoEUD-IQ4_NL397B182.60 GiB0.53 GiB183.73 GiB8.27 GiB23±37%
MiMo-V2.5MoEKV unresolvedQ4_1311B180.99 GiB2.11 GiB183.69 GiB8.31 GiB19±37%
Qwen3-235B-A22BMoEQ6_K235B179.80 GiB3.30 GiB183.69 GiB8.31 GiB15±37%
Qwen3-235B-A22B-Instruct-2507MoEQ6_K235B179.80 GiB3.30 GiB183.69 GiB8.31 GiB15±37%
Qwen3-235B-A22B-Thinking-2507MoEQ6_K235B179.80 GiB3.30 GiB183.69 GiB8.31 GiB15±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ6_K236B179.76 GiB3.30 GiB183.65 GiB8.35 GiB15±37%
Qwen3-VL-235B-A22B-InstructMoEQ6_K236B179.76 GiB3.30 GiB183.65 GiB8.35 GiB15±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q6_K235B179.76 GiB3.30 GiB183.65 GiB8.35 GiB15±37%
grok-2MoEQ5_K_M270B178.41 GiB4.50 GiB183.61 GiB8.39 GiB7±37%
MiMo-V2-FlashMoEKV unresolvedQ4_1310B180.12 GiB2.11 GiB182.83 GiB9.17 GiB20±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ6_K236B180.25 GiB1.19 GiB182.03 GiB9.97 GiB19±37%
DeepSeek-V2.5MoEQ6_K236B180.25 GiB1.19 GiB182.03 GiB9.97 GiB19±37%
DeepSeek-Coder-V2-InstructMoEQ6_K236B180.25 GiB1.19 GiB182.03 GiB9.97 GiB19±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ3_K_M402B177.95 GiB3.38 GiB181.90 GiB10.10 GiB23±37%
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16MoEUD-IQ2_M561B180.64 GiB0.00 GiB181.23 GiB10.77 GiB20±37%
Hy3MoEQ4_1299B174.90 GiB5.63 GiB181.11 GiB10.89 GiB16±37%
DeepSeek-V3.1MoEUD-IQ1_S685B179.11 GiB1.21 GiB180.94 GiB11.06 GiB20±37%
MiniMax-M2.1MoEQ6_K229B174.91 GiB4.36 GiB179.81 GiB12.19 GiB18±37%
MiniMax-M2MoEQ6_K229B174.91 GiB4.36 GiB179.81 GiB12.19 GiB18±37%
MiniMax-M2.5MoEQ6_K229B174.87 GiB4.36 GiB179.76 GiB12.24 GiB18±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ6_K229B174.87 GiB4.36 GiB179.76 GiB12.24 GiB18±37%
ERNIE-4.5-300B-A47B-PTQ4_1300B174.57 GiB3.80 GiB179.04 GiB12.96 GiB4±8.3%
Trinity-Large-PreviewMoEQ3_K_M399B176.43 GiB1.28 GiB178.28 GiB13.72 GiB24±37%
Trinity-Large-TrueBaseMoEI1-Q3_K_M399B176.30 GiB1.28 GiB178.16 GiB13.84 GiB24±37%
Trinity-Large-ThinkingMoEIQ3_M399B176.24 GiB1.28 GiB178.10 GiB13.90 GiB24±37%
Nex-N2-ProMoEIQ3_M397B176.93 GiB0.53 GiB178.06 GiB13.94 GiB24±37%
command-a-plus-05-2026-bf16MoEQ6_K219B176.79 GiB0.68 GiB178.03 GiB13.97 GiB17±37%
GLM-4.7-REAP-218B-A32BMoEQ6_K218B167.57 GiB6.47 GiB174.63 GiB17.37 GiB13±37%
Qwen3.5-REAP-262B-A17BMoEQ5_K_M262B172.96 GiB0.53 GiB174.09 GiB17.91 GiB22±37%
Minimax-M3-abliterated-cleanMoEQ3_K_S427B171.06 GiB2.11 GiB173.74 GiB18.26 GiB20±37%
GLM-5.2MoEQ3_K_M753B169.33 GiB1.54 GiB171.47 GiB20.53 GiB21±37%
Solar-Open2-250BMoEQ5_K_M250B165.65 GiB3.38 GiB169.60 GiB22.40 GiB20±37%
Step-3.7-FlashQ6_K_L201B160.20 GiB7.04 GiB167.82 GiB24.18 GiB4±8.3%
GLM-5.1MoEIQ1_M754B163.93 GiB1.54 GiB166.07 GiB25.93 GiB22±37%
dots.llm1.instMoEQ8_0143B141.37 GiB17.44 GiB159.39 GiB32.61 GiB12±37%
step-3.5-flashQ6_K199B150.81 GiB7.04 GiB158.43 GiB33.57 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 65,536 context with q4_0 KV cache, the largest being GLM-5 at UD-IQ1_S. 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.