Apple · apple

Apple M1 Ultra

Apple M1 Ultra has 128 GB of unified memory at 819 GB/s — about 89.28 GiB usable after driver and compositor overhead. 2089 of 2118 indexed models fit at 16K context with q4_0 KV. Note only 96 GB of its 128 GB is allocatable to the GPU.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
128 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 1794vision language 191audio tts 21image 2audio asr 39video 16embedding 26

What fits at 16K context

largest quantization that fits, per model · 2089 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
command-a-plus-05-2026-bf16MoEIQ3_XS219B95.16 GiB0.26 GiB95.97 GiB0.03 GiB29±37%
ERNIE-4.5-300B-A47B-PTUD-IQ2_XXS300B94.32 GiB0.95 GiB95.94 GiB0.06 GiB7±8.3%
Behemoth-X-123B-v2Q6_K123B93.68 GiB1.55 GiB95.93 GiB0.07 GiB7±8.3%
Mistral-Large-Instruct-2411Q6_K123B93.68 GiB1.55 GiB95.93 GiB0.07 GiB7±8.3%
MiniMax-M2.7MoEUD-Q3_K_M229B94.29 GiB1.09 GiB95.92 GiB0.08 GiB34±37%
Qwen3-235B-A22B-abliteratedMoEI1-IQ3_S235B94.50 GiB0.83 GiB95.91 GiB0.09 GiB28±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ3_K_S236B94.48 GiB0.83 GiB95.89 GiB0.11 GiB28±37%
Qwen3-VL-235B-A22B-InstructMoEQ3_K_S236B94.48 GiB0.83 GiB95.89 GiB0.11 GiB28±37%
Qwen3-235B-A22BMoEQ3_K_S235B94.48 GiB0.83 GiB95.89 GiB0.11 GiB28±37%
MiMo-V2-FlashMoEKV unresolvedUD-IQ2_XXS310B94.58 GiB0.53 GiB95.71 GiB0.29 GiB35±37%
Laguna-S-2.1MoEQ6_K_L118B94.83 GiB0.25 GiB95.66 GiB0.34 GiB34±37%
DeepSeek-Coder-V2-Instruct-0724MoEIQ3_S236B94.70 GiB0.30 GiB95.58 GiB0.42 GiB34±37%
DeepSeek-V2.5MoEIQ3_S236B94.70 GiB0.30 GiB95.58 GiB0.42 GiB34±37%
DeepSeek-Coder-V2-InstructMoEIQ3_S236B94.70 GiB0.30 GiB95.58 GiB0.42 GiB34±37%
grok-2MoEQ2_K_L270B93.41 GiB1.13 GiB95.23 GiB0.77 GiB13±37%
Step-3.5-Flash-REAP-121B-A11BI1-Q6_K121B92.51 GiB1.98 GiB95.07 GiB0.93 GiB7±8.3%
MiniMax-M3MoEIQ1_M427B93.82 GiB0.53 GiB94.91 GiB1.09 GiB35±37%
dots.llm1.instMoEQ4_K_L143B89.95 GiB4.36 GiB94.89 GiB1.11 GiB25±37%
gemma-4-26B-A4B-it-Uncensored-MAXMoEF3225.8B94.02 GiB0.26 GiB94.81 GiB1.19 GiB7±8.3%
MiniMax-M2.1MoEI1-IQ3_M229B93.13 GiB1.09 GiB94.75 GiB1.25 GiB34±37%
MiniMax-M2.5MoEI1-IQ3_M229B93.13 GiB1.09 GiB94.75 GiB1.25 GiB34±37%
Mixtral-8x22B-Instruct-v0.1MoEQ5_K_M141B93.11 GiB0.98 GiB94.70 GiB1.30 GiB13±37%
Mixtral-8x22B-v0.1MoEQ5_K_M141B93.11 GiB0.98 GiB94.70 GiB1.30 GiB13±37%
Mixtral-8x22B-v0.1MoEQ5_K_M141B93.10 GiB0.98 GiB94.70 GiB1.30 GiB13±37%
GLM-4.6-REAP-268B-A32BMoEQ2_K_L269B92.11 GiB1.62 GiB94.32 GiB1.68 GiB26±37%
Qwen3.5-122B-A10B-hereticMoEI1-Q6_K123B93.42 GiB0.11 GiB94.10 GiB1.90 GiB37±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedTQ1_0402B92.59 GiB0.84 GiB94.01 GiB1.99 GiB42±37%
MiniMax-M2MoEQ3_K_S229B92.31 GiB1.09 GiB93.93 GiB2.07 GiB34±37%
GLM-4.5-Air-DerestrictedMoEQ6_K110B92.37 GiB0.81 GiB93.76 GiB2.24 GiB28±37%
GLM-4.5-AirMoEQ6_K110B92.37 GiB0.81 GiB93.76 GiB2.24 GiB28±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q5_K_M139B91.98 GiB1.09 GiB93.61 GiB2.39 GiB30±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-Q5_K_M139B91.98 GiB1.09 GiB93.61 GiB2.39 GiB30±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEI1-IQ3_S229B91.94 GiB1.09 GiB93.56 GiB2.44 GiB34±37%
DeepSeek-V4-FlashMoEQ2_K291B92.86 GiB0.02 GiB93.47 GiB2.53 GiB40±37%
Mistral-Small-4-119B-2603MoEUD-Q6_K119B92.60 GiB0.10 GiB93.28 GiB2.72 GiB37±37%
Step-3.7-FlashUD-IQ4_NL201B90.63 GiB1.98 GiB93.19 GiB2.81 GiB7±8.3%
GLM-4.7MoEUD-IQ1_S358B90.50 GiB1.62 GiB92.71 GiB3.29 GiB29±37%
GLM-4.7-REAP-218B-A32BMoEQ3_K_S218B90.39 GiB1.62 GiB92.59 GiB3.41 GiB24±37%
GLM-4.5MoEUD-IQ1_S358B90.38 GiB1.62 GiB92.59 GiB3.41 GiB29±37%
GLM-4.6MoEUD-IQ1_S357B90.28 GiB1.62 GiB92.49 GiB3.51 GiB29±37%
Qwen3.5-397B-A17BMoEIQ1_M403B91.53 GiB0.13 GiB92.26 GiB3.74 GiB41±37%
Kimi-Linear-48B-A3B-InstructMoEBF1649.1B91.54 GiB0.13 GiB92.23 GiB3.77 GiB7±8.3%
Qwen3-235B-A22B-Instruct-2507MoEIQ3_XS235B90.25 GiB0.83 GiB91.65 GiB4.35 GiB29±37%
Qwen3-235B-A22B-Thinking-2507MoEIQ3_XS235B90.25 GiB0.83 GiB91.65 GiB4.35 GiB29±37%
Solar-Open2-250BMoEIQ3_XXS250B89.86 GiB0.84 GiB91.28 GiB4.72 GiB37±37%
Hermes-4-405BUD-IQ1_M406B88.23 GiB2.21 GiB91.27 GiB4.73 GiB7±8.3%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ5_K_L124B90.28 GiB0.39 GiB91.21 GiB4.79 GiB34±37%
MiMo-V2.5MoEKV unresolvedUD-IQ2_M311B89.93 GiB0.53 GiB91.05 GiB4.95 GiB37±37%
GLM-4.6VMoEQ6_K_L108B89.58 GiB0.81 GiB90.97 GiB5.03 GiB29±37%
Nex-N2-ProMoEIQ1_M397B90.02 GiB0.13 GiB90.75 GiB5.25 GiB42±37%
GLM-4.5VMoEI1-Q6_K108B89.28 GiB0.81 GiB90.66 GiB5.34 GiB29±37%
Hermes-3-Llama-3.1-405BIQ1_M406B87.08 GiB2.21 GiB90.12 GiB5.88 GiB8±8.3%
Trinity-Large-PreviewMoEIQ2_XXS399B88.58 GiB0.49 GiB89.64 GiB6.36 GiB42±37%
Trinity-Large-TrueBaseMoEIQ2_XXS399B88.58 GiB0.49 GiB89.64 GiB6.36 GiB42±37%
gpt-oss-20b-hereticMoEQ5_120.9B88.57 GiB0.11 GiB89.22 GiB6.78 GiB22±37%
Huihui-gpt-oss-20b-BF16-abliteratedMoEQ5_120.9B88.57 GiB0.11 GiB89.22 GiB6.78 GiB22±37%
Dolphin3.0-R1-Mistral-24BF3223.6B87.82 GiB0.70 GiB89.19 GiB6.81 GiB8±8.3%
Dolphin3.0-Mistral-24BF3223.6B87.82 GiB0.70 GiB89.19 GiB6.81 GiB8±8.3%
Mistral-Small-24B-Instruct-2501-abliteratedF3223.6B87.82 GiB0.70 GiB89.19 GiB6.81 GiB8±8.3%
Mistral-Small-24B-Instruct-2501F3223.6B87.82 GiB0.70 GiB89.19 GiB6.81 GiB8±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 M1 Ultra run?
2089 of 2118 indexed open-weight models fit a Apple M1 Ultra at 16,384 context with q4_0 KV cache, the largest being command-a-plus-05-2026-bf16 at IQ3_XS. That covers text, vision-language, image, video and speech models.
How much usable memory does a Apple M1 Ultra actually have?
Its nameplate is 128 GB, but about 89.28 GiB is available to a model once driver and compositor overhead is accounted for, and only 96 GB of the pool can be allocated to the GPU at all.
Is a Apple M1 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.