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. 2082 of 2118 indexed models fit at 128K 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
vision language 190text 1788audio tts 21image 2audio asr 39video 16embedding 26

What fits at 128K context

largest quantization that fits, per model · 2082 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Step-3.7-FlashQ3_K_S201B81.55 GiB13.79 GiB95.92 GiB0.08 GiB7±8.3%
GLM-4.7-REAP-218B-A32BMoEIQ3_XXS218B82.27 GiB12.94 GiB95.79 GiB0.21 GiB15±37%
GLM-4.6-Derestricted-v3MoEIQ2_XXS357B82.26 GiB12.94 GiB95.79 GiB0.21 GiB17±37%
GLM-4.6MoEIQ2_XXS357B82.26 GiB12.94 GiB95.79 GiB0.21 GiB17±37%
HarmonicHarlequin_v5-20BI1-Q5_K_M33.3B21.98 GiB73.13 GiB95.70 GiB0.30 GiB7±8.3%
Qwen3-VL-235B-A22B-ThinkingMoEUD-IQ3_XXS236B88.22 GiB6.61 GiB95.42 GiB0.58 GiB21±37%
Qwen3-VL-235B-A22B-InstructMoEUD-IQ3_XXS236B88.15 GiB6.61 GiB95.34 GiB0.66 GiB21±37%
Devstral-2-123B-Instruct-2512Q5_K_M125B82.25 GiB12.38 GiB95.33 GiB0.67 GiB7±8.3%
Mistral-Medium-3.5-128BI1-Q5_K_M128B82.25 GiB12.38 GiB95.33 GiB0.67 GiB7±8.3%
XORTRON-NXTXPRTXXLI1-Q5_K_M128B82.25 GiB12.38 GiB95.33 GiB0.67 GiB7±8.3%
ERNIE-4.5-300B-A47B-PTUD-IQ1_M300B86.78 GiB7.59 GiB95.05 GiB0.95 GiB7±8.3%
MiMo-V2-FlashMoEKV unresolvedIQ2_M310B90.09 GiB4.22 GiB94.91 GiB1.09 GiB27±37%
Qwen3.5-122B-A10B-hereticMoEI1-Q6_K123B93.42 GiB0.84 GiB94.84 GiB1.16 GiB35±37%
MiMo-V2.5MoEKV unresolvedUD-IQ2_M311B89.93 GiB4.22 GiB94.74 GiB1.26 GiB27±37%
step-3.5-flashQ3_K_S199B80.06 GiB13.79 GiB94.43 GiB1.57 GiB7±8.3%
Step-3.5-Flash-REAP-121B-A11BI1-Q5_K_M121B79.79 GiB13.79 GiB94.16 GiB1.84 GiB7±8.3%
Dolphin3.0-R1-Mistral-24BF3223.6B87.82 GiB5.63 GiB94.11 GiB1.89 GiB7±8.3%
Dolphin3.0-Mistral-24BF3223.6B87.82 GiB5.63 GiB94.11 GiB1.89 GiB7±8.3%
Mistral-Small-24B-Instruct-2501-abliteratedF3223.6B87.82 GiB5.63 GiB94.11 GiB1.89 GiB7±8.3%
Mistral-Small-24B-Instruct-2501F3223.6B87.82 GiB5.63 GiB94.11 GiB1.89 GiB7±8.3%
Mistral-Small-4-119B-2603MoEUD-Q6_K119B92.60 GiB0.79 GiB93.97 GiB2.03 GiB35±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ5_K_L124B90.28 GiB3.09 GiB93.92 GiB2.08 GiB28±37%
Qwen3-235B-A22B-Thinking-2507MoEIQ3_XXS235B86.68 GiB6.61 GiB93.87 GiB2.13 GiB21±37%
Qwen3-235B-A22B-Instruct-2507MoEIQ3_XXS235B86.68 GiB6.61 GiB93.87 GiB2.13 GiB21±37%
Behemoth-X-123B-v2Q5_K_M123B80.55 GiB12.38 GiB93.63 GiB2.37 GiB7±8.3%
Mistral-Large-Instruct-2411Q5_K_M123B80.55 GiB12.38 GiB93.63 GiB2.37 GiB7±8.3%
Laguna-S-2.1MoEUD-Q6_K118B91.19 GiB1.73 GiB93.49 GiB2.51 GiB31±37%
DeepSeek-V4-FlashMoEQ2_K291B92.86 GiB0.02 GiB93.47 GiB2.53 GiB40±37%
Qwen3.5-397B-A17BMoEIQ1_M403B91.53 GiB1.05 GiB93.19 GiB2.81 GiB38±37%
MiniMax-M2.1MoEIQ3_XXS229B83.91 GiB8.72 GiB93.17 GiB2.83 GiB22±37%
MiniMax-M2MoEIQ3_XXS229B83.91 GiB8.72 GiB93.17 GiB2.83 GiB22±37%
Kimi-Linear-48B-A3B-InstructMoEBF1649.1B91.54 GiB1.07 GiB93.16 GiB2.84 GiB7±8.3%
command-a-plus-05-2026-bf16MoEQ3_K_S219B91.28 GiB1.24 GiB93.08 GiB2.92 GiB28±37%
DeepSeek-Coder-V2-Instruct-0724MoEIQ3_XS236B89.69 GiB2.37 GiB92.65 GiB3.35 GiB30±37%
DeepSeek-V2.5MoEIQ3_XS236B89.69 GiB2.37 GiB92.65 GiB3.35 GiB30±37%
DeepSeek-Coder-V2-InstructMoEIQ3_XS236B89.69 GiB2.37 GiB92.65 GiB3.35 GiB30±37%
grok-2MoEUD-IQ1_S270B82.82 GiB9.00 GiB92.52 GiB3.48 GiB11±37%
GLM-4.7MoEUD-TQ1_0358B78.69 GiB12.94 GiB92.22 GiB3.78 GiB17±37%
dots.llm1.instMoEQ2_K_L143B56.67 GiB34.88 GiB92.12 GiB3.88 GiB9±37%
xLAM-8x7b-rMoEBF1646.7B86.99 GiB4.50 GiB92.08 GiB3.92 GiB12±37%
GLM-4.5MoEUD-TQ1_0358B78.54 GiB12.94 GiB92.07 GiB3.93 GiB17±37%
Nex-N2-ProMoEIQ1_M397B90.02 GiB1.05 GiB91.67 GiB4.33 GiB38±37%
Trinity-Large-PreviewMoEIQ2_XXS399B88.58 GiB2.33 GiB91.48 GiB4.52 GiB35±37%
Trinity-Large-TrueBaseMoEIQ2_XXS399B88.58 GiB2.33 GiB91.48 GiB4.52 GiB35±37%
Codestral-22B-v0.1F3222.2B82.88 GiB7.88 GiB91.36 GiB4.64 GiB7±8.3%
Qwen3-235B-A22B-abliteratedMoEI1-IQ3_XXS235B84.17 GiB6.61 GiB91.36 GiB4.64 GiB21±37%
Mixtral-8x22B-v0.1MoEF32141B82.84 GiB7.88 GiB91.33 GiB4.67 GiB12±37%
GLM-4.5-Air-REAP-82B-A12BMoEQ8_081.9B84.17 GiB6.47 GiB91.21 GiB4.79 GiB20±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEI1-IQ3_XXS229B81.91 GiB8.72 GiB91.17 GiB4.83 GiB22±37%
MiniMax-M2.5MoEI1-IQ3_XXS229B81.91 GiB8.72 GiB91.17 GiB4.83 GiB22±37%
Gemma-4-Dark-Gemistry-31BQ6_K32.7B84.44 GiB5.95 GiB91.02 GiB4.98 GiB7±8.3%
Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTPQ6_K27.8B88.00 GiB2.25 GiB90.87 GiB5.13 GiB7±8.3%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q4_1139B81.21 GiB8.72 GiB90.46 GiB5.54 GiB20±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-Q4_1139B81.21 GiB8.72 GiB90.46 GiB5.54 GiB20±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ6_K_L109B83.13 GiB6.75 GiB90.46 GiB5.54 GiB21±37%
GLM-4.6-REAP-268B-A32BMoEUD-IQ1_M269B76.78 GiB12.94 GiB90.31 GiB5.69 GiB16±37%
gpt-oss-20b-hereticMoEQ5_120.9B88.57 GiB0.85 GiB89.95 GiB6.05 GiB21±37%
Huihui-gpt-oss-20b-BF16-abliteratedMoEQ5_120.9B88.57 GiB0.85 GiB89.95 GiB6.05 GiB21±37%
CalmeRys-78B-Orpo-v0.1Q8_078.0B77.16 GiB12.09 GiB89.93 GiB6.07 GiB8±8.3%
calme-2.3-rys-78bQ8_078.0B77.16 GiB12.09 GiB89.93 GiB6.07 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?
2082 of 2118 indexed open-weight models fit a Apple M1 Ultra at 131,072 context with q4_0 KV cache, the largest being Step-3.7-Flash at Q3_K_S. 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.