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 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 64K context

largest quantization that fits, per model · 2108 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
GLM-4.6-Derestricted-v3MoEIQ4_XS357B178.83 GiB12.22 GiB191.64 GiB0.36 GiB12±37%
GLM-4.6MoEIQ4_XS357B178.83 GiB12.22 GiB191.64 GiB0.36 GiB12±37%
GLM-4.6-REAP-268B-A32BMoEQ5_K_M269B177.71 GiB12.22 GiB190.52 GiB1.48 GiB11±37%
openPangu-2.0-FlashMoEKV unresolvedBF16100B188.04 GiB1.72 GiB190.31 GiB1.69 GiB19±37%
cogito-671b-v2.1MoEUD-IQ1_M671B187.19 GiB2.28 GiB190.10 GiB1.90 GiB19±37%
DeepSeek-V3.1-TerminusMoEUD-IQ1_M685B187.19 GiB2.28 GiB190.09 GiB1.91 GiB19±37%
DeepSeek-V3-0324MoEUD-IQ1_M685B186.94 GiB2.28 GiB189.85 GiB2.15 GiB19±37%
cogito-v2-preview-deepseek-671B-MoEMoEUD-IQ1_M671B186.85 GiB2.28 GiB189.76 GiB2.24 GiB19±37%
DeepSeek-R1-0528MoEUD-IQ1_M685B186.69 GiB2.28 GiB189.60 GiB2.40 GiB19±37%
DeepSeek-TNG-R1T2-ChimeraMoEUD-IQ1_M685B186.69 GiB2.28 GiB189.60 GiB2.40 GiB19±37%
DeepSeek-Prover-V2-671BMoEUD-IQ1_M685B186.06 GiB2.28 GiB188.97 GiB3.03 GiB19±37%
DeepSeek-V3.2MoEUD-IQ1_M685B185.62 GiB2.28 GiB188.53 GiB3.47 GiB19±37%
MiniMax-M3MoEQ3_K_M427B183.11 GiB3.98 GiB187.66 GiB4.34 GiB18±37%
grok-2MoEQ5_K_M270B178.41 GiB8.50 GiB187.61 GiB4.39 GiB6±37%
Hermes-3-Llama-3.1-405BIQ3_M406B169.26 GiB16.73 GiB186.83 GiB5.17 GiB4±8.3%
Qwen3.5-397B-A17BMoEUD-IQ4_NL403B185.08 GiB1.00 GiB186.68 GiB5.32 GiB22±37%
Qwen3-235B-A22BMoEQ6_K235B179.80 GiB6.24 GiB186.63 GiB5.37 GiB13±37%
Qwen3-235B-A22B-Instruct-2507MoEQ6_K235B179.80 GiB6.24 GiB186.63 GiB5.37 GiB13±37%
Qwen3-235B-A22B-Thinking-2507MoEQ6_K235B179.80 GiB6.24 GiB186.63 GiB5.37 GiB13±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ6_K236B179.76 GiB6.24 GiB186.58 GiB5.42 GiB13±37%
Qwen3-VL-235B-A22B-InstructMoEQ6_K236B179.76 GiB6.24 GiB186.58 GiB5.42 GiB13±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q6_K235B179.76 GiB6.24 GiB186.58 GiB5.42 GiB13±37%
r1-1776MoEIQ2_S671B183.47 GiB2.28 GiB186.37 GiB5.63 GiB19±37%
DeepSeek-R1MoEIQ2_S685B183.47 GiB2.28 GiB186.37 GiB5.63 GiB19±37%
Hy3MoEQ4_1299B174.90 GiB10.63 GiB186.11 GiB5.89 GiB13±37%
MiMo-V2.5MoEKV unresolvedQ4_1311B180.99 GiB3.98 GiB185.57 GiB6.43 GiB18±37%
Qwen3-Coder-480B-A35B-InstructMoEIQ3_XXS480B176.37 GiB8.23 GiB185.19 GiB6.81 GiB14±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ3_K_M402B177.95 GiB6.38 GiB184.90 GiB7.10 GiB19±37%
MiMo-V2-FlashMoEKV unresolvedQ4_1310B180.12 GiB3.98 GiB184.70 GiB7.30 GiB18±37%
Ornith-1.0-397BMoEUD-IQ4_NL397B182.60 GiB1.00 GiB184.19 GiB7.81 GiB23±37%
MiniMax-M2.7MoEUD-Q6_K229B175.15 GiB8.23 GiB183.91 GiB8.09 GiB15±37%
MiniMax-M2.1MoEQ6_K229B174.91 GiB8.23 GiB183.68 GiB8.32 GiB15±37%
MiniMax-M2MoEQ6_K229B174.91 GiB8.23 GiB183.68 GiB8.32 GiB15±37%
MiniMax-M2.5MoEQ6_K229B174.87 GiB8.23 GiB183.64 GiB8.36 GiB15±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ6_K229B174.87 GiB8.23 GiB183.64 GiB8.36 GiB15±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ6_K236B180.25 GiB2.24 GiB183.08 GiB8.92 GiB18±37%
DeepSeek-V2.5MoEQ6_K236B180.25 GiB2.24 GiB183.08 GiB8.92 GiB18±37%
DeepSeek-Coder-V2-InstructMoEQ6_K236B180.25 GiB2.24 GiB183.08 GiB8.92 GiB18±37%
ERNIE-4.5-300B-A47B-PTQ4_1300B174.57 GiB7.17 GiB182.41 GiB9.59 GiB4±8.3%
DeepSeek-V3.1MoEUD-IQ1_S685B179.11 GiB2.28 GiB182.01 GiB9.99 GiB19±37%
Llama-3_1-Nemotron-51B-InstructF1651.5B95.94 GiB85.00 GiB181.63 GiB10.37 GiB4±8.3%
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16MoEUD-IQ2_M561B180.64 GiB0.00 GiB181.23 GiB10.77 GiB20±37%
GLM-4.7-REAP-218B-A32BMoEQ6_K218B167.57 GiB12.22 GiB180.38 GiB11.62 GiB11±37%
Trinity-Large-PreviewMoEQ3_K_M399B176.43 GiB2.41 GiB179.42 GiB12.58 GiB22±37%
Trinity-Large-TrueBaseMoEI1-Q3_K_M399B176.30 GiB2.41 GiB179.29 GiB12.71 GiB22±37%
Trinity-Large-ThinkingMoEIQ3_M399B176.24 GiB2.41 GiB179.23 GiB12.77 GiB22±37%
command-a-plus-05-2026-bf16MoEQ6_K219B176.79 GiB1.29 GiB178.63 GiB13.37 GiB16±37%
Llama-3_3-Nemotron-Super-49B-v1_5BF1649.9B92.89 GiB85.00 GiB178.59 GiB13.41 GiB4±8.3%
Valkyrie-49B-v2.1BF1649.9B92.89 GiB85.00 GiB178.59 GiB13.41 GiB4±8.3%
Llama-3_3-Nemotron-Super-49B-v1BF1649.9B92.89 GiB85.00 GiB178.59 GiB13.41 GiB4±8.3%
Nex-N2-ProMoEIQ3_M397B176.93 GiB1.00 GiB178.53 GiB13.47 GiB23±37%
Minimax-M3-abliterated-cleanMoEQ3_K_S427B171.06 GiB3.98 GiB175.61 GiB16.39 GiB19±37%
dots.llm1.instMoEQ8_0143B141.37 GiB32.94 GiB174.89 GiB17.11 GiB8±37%
Qwen3.5-REAP-262B-A17BMoEQ5_K_M262B172.96 GiB1.00 GiB174.56 GiB17.44 GiB22±37%
Step-3.7-FlashQ6_K_L201B160.20 GiB13.30 GiB174.08 GiB17.92 GiB4±8.3%
GLM-5.2MoEQ3_K_M753B169.33 GiB2.91 GiB172.84 GiB19.16 GiB20±37%
Solar-Open2-250BMoEQ5_K_M250B165.65 GiB6.38 GiB172.60 GiB19.40 GiB18±37%
GLM-4.5MoEQ3_K_M358B159.51 GiB12.22 GiB172.32 GiB19.68 GiB13±37%
GLM-4.7MoEQ3_K_M358B159.51 GiB12.22 GiB172.32 GiB19.68 GiB13±37%
Qwen3-Coder-REAP-363B-A35BMoEQ3_K_M363B161.57 GiB8.23 GiB170.39 GiB21.61 GiB13±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 65,536 context with q8_0 KV cache, the largest being GLM-4.6-Derestricted-v3 at IQ4_XS. 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.