Apple · apple

Apple M5 Max

Apple M5 Max has 128 GB of unified memory at 614 GB/s — about 89.28 GiB usable after driver and compositor overhead. 2087 of 2118 indexed models fit at 64K 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
LPDDR5X-9600
Bandwidth
614 GB/s
512-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 1793vision language 190audio tts 21image 2audio asr 39video 16embedding 26

What fits at 64K context

largest quantization that fits, per model · 2087 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
dots.llm1.instMoEQ4_0143B77.44 GiB17.44 GiB95.45 GiB0.55 GiB11±37%
gemma-4-26B-A4B-it-Uncensored-MAXMoEF3225.8B94.02 GiB0.79 GiB95.34 GiB0.66 GiB5±8.3%
Llama-3_3-Nemotron-Super-49B-v1_5Q8_049.9B49.36 GiB45.00 GiB95.05 GiB0.95 GiB5±8.3%
Valkyrie-49B-v2.1Q8_049.9B49.36 GiB45.00 GiB95.05 GiB0.95 GiB5±8.3%
Llama-3_3-Nemotron-Super-49B-v1Q8_049.9B49.36 GiB45.00 GiB95.05 GiB0.95 GiB5±8.3%
Step-3.7-FlashQ3_K_M201B87.36 GiB7.04 GiB94.98 GiB1.02 GiB5±8.3%
Mixtral-8x22B-Instruct-v0.1MoEQ5_K_S141B90.32 GiB3.94 GiB94.87 GiB1.13 GiB9±37%
Mixtral-8x22B-v0.1MoEQ5_K_S141B90.31 GiB3.94 GiB94.86 GiB1.14 GiB9±37%
Mixtral-8x22B-v0.1MoEQ5_K_S141B90.31 GiB3.94 GiB94.86 GiB1.14 GiB9±37%
Qwen3.5-122B-A10B-hereticMoEI1-Q6_K123B93.42 GiB0.42 GiB94.42 GiB1.58 GiB28±37%
GLM-4.6-REAP-268B-A32BMoEUD-IQ2_M269B87.26 GiB6.47 GiB94.32 GiB1.68 GiB16±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q5_K_S139B89.27 GiB4.36 GiB94.17 GiB1.83 GiB19±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-Q5_K_S139B89.27 GiB4.36 GiB94.17 GiB1.83 GiB19±37%
Qwen3-235B-A22B-Instruct-2507MoEIQ3_XS235B90.25 GiB3.30 GiB94.13 GiB1.87 GiB19±37%
Qwen3-235B-A22B-Thinking-2507MoEIQ3_XS235B90.25 GiB3.30 GiB94.13 GiB1.87 GiB19±37%
Solar-Open2-250BMoEIQ3_XXS250B89.86 GiB3.38 GiB93.81 GiB2.19 GiB24±37%
Mistral-Small-4-119B-2603MoEUD-Q6_K119B92.60 GiB0.40 GiB93.58 GiB2.42 GiB29±37%
MiniMax-M2.7MoEIQ3_XXS229B88.65 GiB4.36 GiB93.54 GiB2.46 GiB21±37%
DeepSeek-V4-FlashMoEQ2_K291B92.86 GiB0.02 GiB93.47 GiB2.53 GiB32±37%
GLM-4.6VMoEQ6_K_L108B89.58 GiB3.23 GiB93.40 GiB2.60 GiB19±37%
Qwen3-235B-A22B-abliteratedMoEI1-IQ3_XS235B89.36 GiB3.30 GiB93.25 GiB2.75 GiB19±37%
grok-2MoEUD-IQ1_M270B88.02 GiB4.50 GiB93.21 GiB2.79 GiB9±37%
GLM-4.5VMoEI1-Q6_K108B89.28 GiB3.23 GiB93.09 GiB2.91 GiB19±37%
MiMo-V2-FlashMoEKV unresolvedIQ2_M310B90.09 GiB2.11 GiB92.80 GiB3.20 GiB25±37%
Mistral-Medium-3.5-128BQ5_K_L128B85.78 GiB6.19 GiB92.67 GiB3.33 GiB6±8.3%
Qwen3.5-397B-A17BMoEIQ1_M403B91.53 GiB0.53 GiB92.66 GiB3.34 GiB32±37%
Laguna-S-2.1MoEUD-Q6_K118B91.19 GiB0.88 GiB92.65 GiB3.35 GiB26±37%
Hy3MoEIQ2_S299B86.43 GiB5.63 GiB92.64 GiB3.36 GiB19±37%
MiMo-V2.5MoEKV unresolvedUD-IQ2_M311B89.93 GiB2.11 GiB92.63 GiB3.37 GiB25±37%
Kimi-Linear-48B-A3B-InstructMoEBF1649.1B91.54 GiB0.53 GiB92.63 GiB3.37 GiB6±8.3%
GLM-4.7-REAP-218B-A32BMoEIQ3_XS218B85.49 GiB6.47 GiB92.55 GiB3.45 GiB15±37%
command-a-plus-05-2026-bf16MoEQ3_K_S219B91.28 GiB0.68 GiB92.52 GiB3.48 GiB23±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ5_K_L124B90.28 GiB1.55 GiB92.37 GiB3.63 GiB24±37%
MiniMax-M2.1MoEIQ3_XS229B87.32 GiB4.36 GiB92.22 GiB3.78 GiB22±37%
MiniMax-M2MoEIQ3_XS229B87.32 GiB4.36 GiB92.22 GiB3.78 GiB22±37%
Qwen3-VL-235B-A22B-ThinkingMoEUD-IQ3_XXS236B88.22 GiB3.30 GiB92.11 GiB3.89 GiB19±37%
Qwen3-VL-235B-A22B-InstructMoEUD-IQ3_XXS236B88.15 GiB3.30 GiB92.04 GiB3.96 GiB19±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEI1-IQ3_XS229B86.92 GiB4.36 GiB91.81 GiB4.19 GiB22±37%
MiniMax-M2.5MoEI1-IQ3_XS229B86.92 GiB4.36 GiB91.81 GiB4.19 GiB22±37%
step-3.5-flashQ3_K_M199B84.04 GiB7.04 GiB91.65 GiB4.35 GiB6±8.3%
DeepSeek-Coder-V2-Instruct-0724MoEIQ3_XS236B89.69 GiB1.19 GiB91.47 GiB4.53 GiB26±37%
DeepSeek-V2.5MoEIQ3_XS236B89.69 GiB1.19 GiB91.47 GiB4.53 GiB26±37%
DeepSeek-Coder-V2-InstructMoEIQ3_XS236B89.69 GiB1.19 GiB91.47 GiB4.53 GiB26±37%
Dolphin3.0-R1-Mistral-24BF3223.6B87.82 GiB2.81 GiB91.30 GiB4.70 GiB6±8.3%
Dolphin3.0-Mistral-24BF3223.6B87.82 GiB2.81 GiB91.30 GiB4.70 GiB6±8.3%
Mistral-Small-24B-Instruct-2501-abliteratedF3223.6B87.82 GiB2.81 GiB91.30 GiB4.70 GiB6±8.3%
Mistral-Small-24B-Instruct-2501F3223.6B87.82 GiB2.81 GiB91.30 GiB4.70 GiB6±8.3%
ERNIE-4.5-300B-A47B-PTUD-IQ1_M300B86.78 GiB3.80 GiB91.25 GiB4.75 GiB6±8.3%
Nex-N2-ProMoEIQ1_M397B90.02 GiB0.53 GiB91.15 GiB4.85 GiB32±37%
Hermes-4-405BUD-IQ1_S406B81.23 GiB8.86 GiB90.92 GiB5.08 GiB6±8.3%
Trinity-Large-PreviewMoEIQ2_XXS399B88.58 GiB1.28 GiB90.43 GiB5.57 GiB31±37%
Trinity-Large-TrueBaseMoEIQ2_XXS399B88.58 GiB1.28 GiB90.43 GiB5.57 GiB31±37%
GLM-4.5MoEIQ2_XXS358B83.01 GiB6.47 GiB90.07 GiB5.93 GiB18±37%
xLAM-8x7b-rMoEBF1646.7B86.99 GiB2.25 GiB89.83 GiB6.17 GiB10±37%
GLM-4.7MoEIQ2_XXS358B82.69 GiB6.47 GiB89.75 GiB6.25 GiB18±37%
Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTPQ6_K27.8B88.00 GiB1.13 GiB89.74 GiB6.26 GiB6±8.3%
gpt-oss-20b-hereticMoEQ5_120.9B88.57 GiB0.43 GiB89.53 GiB6.47 GiB17±37%
Huihui-gpt-oss-20b-BF16-abliteratedMoEQ5_120.9B88.57 GiB0.43 GiB89.53 GiB6.47 GiB17±37%
GLM-4.6-Derestricted-v3MoEIQ2_XXS357B82.26 GiB6.47 GiB89.32 GiB6.68 GiB18±37%
GLM-4.6MoEIQ2_XXS357B82.26 GiB6.47 GiB89.32 GiB6.68 GiB18±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 M5 Max run?
2087 of 2118 indexed open-weight models fit a Apple M5 Max at 65,536 context with q4_0 KV cache, the largest being dots.llm1.inst at Q4_0. That covers text, vision-language, image, video and speech models.
How much usable memory does a Apple M5 Max 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 M5 Max fast for local AI?
Its memory bandwidth is 614 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.