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. 2084 of 2118 indexed models fit at 64K context with q8_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
vision language 190text 1790audio tts 21image 2audio asr 39video 16embedding 26

What fits at 64K context

largest quantization that fits, per model · 2084 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
GLM-4.6VMoEQ6_K108B89.30 GiB6.11 GiB95.99 GiB0.01 GiB21±37%
MiniMax-M2.7MoEUD-Q3_K_S229B87.21 GiB8.23 GiB95.98 GiB0.02 GiB22±37%
Valkyrie-49B-v2.1I1-IQ1_S49.9B10.27 GiB85.00 GiB95.96 GiB0.04 GiB7±8.3%
Llama-3_3-Nemotron-Super-49B-v1IQ1_S49.9B10.27 GiB85.00 GiB95.96 GiB0.04 GiB7±8.3%
GLM-4.5VMoEI1-Q6_K108B89.28 GiB6.11 GiB95.96 GiB0.04 GiB21±37%
MiniMax-M2MoEUD-IQ3_XXS229B87.18 GiB8.23 GiB95.95 GiB0.05 GiB22±37%
GLM-4.6-REAP-268B-A32BMoEUD-IQ2_XXS269B83.03 GiB12.22 GiB95.84 GiB0.16 GiB16±37%
Solar-Open2-250BMoEQ2_K250B88.87 GiB6.38 GiB95.82 GiB0.18 GiB25±37%
GLM-4.5MoEIQ2_XXS358B83.01 GiB12.22 GiB95.82 GiB0.18 GiB17±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEI1-IQ3_XS229B86.92 GiB8.23 GiB95.69 GiB0.31 GiB22±37%
MiniMax-M2.1MoEI1-IQ3_XS229B86.92 GiB8.23 GiB95.69 GiB0.31 GiB22±37%
MiniMax-M2.5MoEI1-IQ3_XS229B86.92 GiB8.23 GiB95.69 GiB0.31 GiB22±37%
GLM-4.7MoEIQ2_XXS358B82.69 GiB12.22 GiB95.50 GiB0.50 GiB17±37%
Step-3.7-FlashQ3_K_S201B81.55 GiB13.30 GiB95.43 GiB0.57 GiB7±8.3%
dots.llm1.instMoEIQ3_XS143B61.71 GiB32.94 GiB95.22 GiB0.78 GiB10±37%
HarmonicHarlequin_v5-20BI1-Q6_K33.3B25.46 GiB69.06 GiB95.11 GiB0.89 GiB7±8.3%
GLM-4.7-REAP-218B-A32BMoEIQ3_XXS218B82.27 GiB12.22 GiB95.08 GiB0.92 GiB15±37%
GLM-4.6-Derestricted-v3MoEIQ2_XXS357B82.26 GiB12.22 GiB95.07 GiB0.93 GiB17±37%
GLM-4.6MoEIQ2_XXS357B82.26 GiB12.22 GiB95.07 GiB0.93 GiB17±37%
Qwen3-VL-235B-A22B-ThinkingMoEUD-IQ3_XXS236B88.22 GiB6.24 GiB95.05 GiB0.95 GiB21±37%
Qwen3-VL-235B-A22B-InstructMoEUD-IQ3_XXS236B88.15 GiB6.24 GiB94.97 GiB1.03 GiB21±37%
Qwen3.5-122B-A10B-hereticMoEI1-Q6_K123B93.42 GiB0.80 GiB94.79 GiB1.21 GiB35±37%
MiMo-V2-FlashMoEKV unresolvedIQ2_M310B90.09 GiB3.98 GiB94.68 GiB1.32 GiB28±37%
Devstral-2-123B-Instruct-2512Q5_K_M125B82.25 GiB11.69 GiB94.65 GiB1.35 GiB7±8.3%
Mistral-Medium-3.5-128BI1-Q5_K_M128B82.25 GiB11.69 GiB94.65 GiB1.35 GiB7±8.3%
XORTRON-NXTXPRTXXLI1-Q5_K_M128B82.25 GiB11.69 GiB94.65 GiB1.35 GiB7±8.3%
ERNIE-4.5-300B-A47B-PTUD-IQ1_M300B86.78 GiB7.17 GiB94.63 GiB1.37 GiB7±8.3%
MiMo-V2.5MoEKV unresolvedUD-IQ2_M311B89.93 GiB3.98 GiB94.51 GiB1.49 GiB28±37%
step-3.5-flashQ3_K_S199B80.06 GiB13.30 GiB93.94 GiB2.06 GiB7±8.3%
Mistral-Small-4-119B-2603MoEUD-Q6_K119B92.60 GiB0.75 GiB93.93 GiB2.07 GiB35±37%
Dolphin3.0-R1-Mistral-24BF3223.6B87.82 GiB5.31 GiB93.80 GiB2.20 GiB7±8.3%
Dolphin3.0-Mistral-24BF3223.6B87.82 GiB5.31 GiB93.80 GiB2.20 GiB7±8.3%
Mistral-Small-24B-Instruct-2501-abliteratedF3223.6B87.82 GiB5.31 GiB93.80 GiB2.20 GiB7±8.3%
Mistral-Small-24B-Instruct-2501F3223.6B87.82 GiB5.31 GiB93.80 GiB2.20 GiB7±8.3%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ5_K_L124B90.28 GiB2.92 GiB93.74 GiB2.26 GiB28±37%
Step-3.5-Flash-REAP-121B-A11BI1-Q5_K_M121B79.79 GiB13.30 GiB93.66 GiB2.34 GiB7±8.3%
Qwen3-235B-A22B-Thinking-2507MoEIQ3_XXS235B86.68 GiB6.24 GiB93.51 GiB2.49 GiB21±37%
Qwen3-235B-A22B-Instruct-2507MoEIQ3_XXS235B86.68 GiB6.24 GiB93.51 GiB2.49 GiB21±37%
DeepSeek-V4-FlashMoEQ2_K291B92.86 GiB0.03 GiB93.49 GiB2.51 GiB40±37%
Laguna-S-2.1MoEUD-Q6_K118B91.19 GiB1.67 GiB93.43 GiB2.57 GiB31±37%
Qwen3.5-397B-A17BMoEIQ1_M403B91.53 GiB1.00 GiB93.13 GiB2.87 GiB38±37%
command-a-plus-05-2026-bf16MoEQ3_K_S219B91.28 GiB1.29 GiB93.12 GiB2.88 GiB28±37%
Kimi-Linear-48B-A3B-InstructMoEBF1649.1B91.54 GiB1.01 GiB93.10 GiB2.90 GiB7±8.3%
Behemoth-X-123B-v2Q5_K_M123B80.55 GiB11.69 GiB92.94 GiB3.06 GiB7±8.3%
Mistral-Large-Instruct-2411Q5_K_M123B80.55 GiB11.69 GiB92.94 GiB3.06 GiB7±8.3%
DeepSeek-Coder-V2-Instruct-0724MoEIQ3_XS236B89.69 GiB2.24 GiB92.52 GiB3.48 GiB30±37%
DeepSeek-V2.5MoEIQ3_XS236B89.69 GiB2.24 GiB92.52 GiB3.48 GiB30±37%
DeepSeek-Coder-V2-InstructMoEIQ3_XS236B89.69 GiB2.24 GiB92.52 GiB3.48 GiB30±37%
grok-2MoEUD-IQ1_S270B82.82 GiB8.50 GiB92.02 GiB3.98 GiB11±37%
xLAM-8x7b-rMoEBF1646.7B86.99 GiB4.25 GiB91.83 GiB4.17 GiB12±37%
Nex-N2-ProMoEIQ1_M397B90.02 GiB1.00 GiB91.62 GiB4.38 GiB38±37%
Trinity-Large-PreviewMoEIQ2_XXS399B88.58 GiB2.41 GiB91.56 GiB4.44 GiB35±37%
Trinity-Large-TrueBaseMoEIQ2_XXS399B88.58 GiB2.41 GiB91.56 GiB4.44 GiB35±37%
Gemma-4-Dark-Gemistry-31BQ6_K32.7B84.44 GiB5.94 GiB91.01 GiB4.99 GiB7±8.3%
Qwen3-235B-A22B-abliteratedMoEI1-IQ3_XXS235B84.17 GiB6.24 GiB90.99 GiB5.01 GiB22±37%
Codestral-22B-v0.1F3222.2B82.88 GiB7.44 GiB90.93 GiB5.07 GiB7±8.3%
Mixtral-8x22B-v0.1MoEF32141B82.84 GiB7.44 GiB90.89 GiB5.11 GiB12±37%
GLM-4.5-Air-REAP-82B-A12BMoEQ8_081.9B84.17 GiB6.11 GiB90.85 GiB5.15 GiB20±37%
Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTPQ6_K27.8B88.00 GiB2.13 GiB90.74 GiB5.26 GiB7±8.3%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ6_K_L109B83.13 GiB6.38 GiB90.08 GiB5.92 GiB22±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 M1 Ultra run?
2084 of 2118 indexed open-weight models fit a Apple M1 Ultra at 65,536 context with q8_0 KV cache, the largest being GLM-4.6V at Q6_K. 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.