Apple · apple

Apple M2 Max

Apple M2 Max has 96 GB of unified memory at 410 GB/s — about 66.96 GiB usable after driver and compositor overhead. 2071 of 2118 indexed models fit at 32K context with q8_0 KV. Note only 72 GB of its 96 GB is allocatable to the GPU.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
96 GB
LPDDR5-6400
Bandwidth
410 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 1781vision language 186image 2audio asr 39audio tts 21video 16embedding 26

What fits at 32K context

largest quantization that fits, per model · 2071 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Mixtral-8x22B-Instruct-v0.1MoEQ3_K_L141B67.60 GiB3.72 GiB71.93 GiB0.07 GiB8±37%
Mixtral-8x22B-v0.1MoEQ3_K_L141B67.60 GiB3.72 GiB71.92 GiB0.08 GiB8±37%
Mixtral-8x22B-v0.1MoEQ3_K_L141B67.60 GiB3.72 GiB71.92 GiB0.08 GiB8±37%
c4ai-command-r-plus-08-2024Q5_K_S104B66.87 GiB4.25 GiB71.85 GiB0.15 GiB5±8.3%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q3_K_L139B67.16 GiB4.12 GiB71.81 GiB0.19 GiB16±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-Q3_K_L139B67.16 GiB4.12 GiB71.81 GiB0.19 GiB16±37%
MiniMax-M2.1MoEIQ2_M229B67.05 GiB4.12 GiB71.70 GiB0.30 GiB18±37%
MiniMax-M2MoEIQ2_M229B67.05 GiB4.12 GiB71.70 GiB0.30 GiB18±37%
OYM-Qimi-122B-A10B-K2.6MoEI1-Q4_K_M125B70.64 GiB0.40 GiB71.61 GiB0.39 GiB25±37%
Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliteratedMoEQ4_K_M123B70.63 GiB0.40 GiB71.61 GiB0.39 GiB25±37%
Behemoth-X-123B-v2Q4_K_S123B64.79 GiB5.84 GiB71.34 GiB0.66 GiB5±8.3%
Mistral-Large-Instruct-2411Q4_K_S123B64.79 GiB5.84 GiB71.34 GiB0.66 GiB5±8.3%
Llama-3_3-Nemotron-Super-49B-v1_5Q4_K_M49.9B28.14 GiB42.50 GiB71.33 GiB0.67 GiB5±8.3%
Valkyrie-49B-v2.1I1-Q4_K_M49.9B28.14 GiB42.50 GiB71.33 GiB0.67 GiB5±8.3%
Llama-3_3-Nemotron-Super-49B-v1Q4_K_M49.9B28.14 GiB42.50 GiB71.33 GiB0.67 GiB5±8.3%
Step-3.5-Flash-REAP-121B-A11BI1-Q4_K_S121B63.83 GiB6.92 GiB71.33 GiB0.67 GiB5±8.3%
Step-3.7-FlashIQ2_M201B63.68 GiB6.92 GiB71.18 GiB0.82 GiB5±8.3%
MiMo-V2-FlashMoEKV unresolvedIQ2_XXS310B68.47 GiB1.99 GiB71.06 GiB0.94 GiB22±37%
Mistral-Medium-3.5-128BIQ4_XS128B64.39 GiB5.84 GiB70.94 GiB1.06 GiB5±8.3%
dots.llm1.instMoEIQ2_M143B53.68 GiB16.47 GiB70.73 GiB1.27 GiB9±37%
CalmeRys-78B-Orpo-v0.1Q6_K78.0B64.27 GiB5.71 GiB70.67 GiB1.33 GiB5±8.3%
calme-2.3-rys-78bQ6_K78.0B64.27 GiB5.71 GiB70.67 GiB1.33 GiB5±8.3%
Llama-3_1-Nemotron-51B-InstructQ4_K_S51.5B27.46 GiB42.50 GiB70.65 GiB1.35 GiB5±8.3%
Qwen3.5-122B-A10BMoEQ4_K_S125B69.66 GiB0.40 GiB70.64 GiB1.36 GiB26±37%
GLM-4.6VMoEQ4_K_L108B66.89 GiB3.05 GiB70.52 GiB1.48 GiB16±37%
Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16MoEQ8_035.1B69.57 GiB0.33 GiB70.46 GiB1.54 GiB26±37%
Laguna-S-2.1MoEQ4_1118B68.96 GiB0.87 GiB70.40 GiB1.60 GiB23±37%
Qwen3.5-122B-A10B-hereticMoEI1-Q4_K_M123B69.11 GiB0.40 GiB70.09 GiB1.91 GiB26±37%
MiniMax-M2.7MoEUD-IQ2_M229B65.32 GiB4.12 GiB69.97 GiB2.03 GiB18±37%
Qwen2.5-Coder-32B-InstructQ8_032.8B64.86 GiB4.25 GiB69.76 GiB2.24 GiB5±8.3%
Mistral-Small-4-119B-2603MoEUD-Q4_K_M119B68.70 GiB0.37 GiB69.65 GiB2.35 GiB26±37%
MiMo-V2.5MoEKV unresolvedIQ1_M311B67.01 GiB1.99 GiB69.60 GiB2.40 GiB22±37%
XORTRON-NXTXPRTXXLIQ4_XS128B63.03 GiB5.84 GiB69.58 GiB2.42 GiB5±8.3%
gpt-oss-120b-Uncensored-xCloudMoEI1-Q4_1117B68.42 GiB0.61 GiB69.57 GiB2.43 GiB25±37%
gpt-oss-120b-abliteratedMoEI1-Q4_1117B68.42 GiB0.61 GiB69.57 GiB2.43 GiB25±37%
HunyuanImage-2.1Q6_K17.5B68.97 GiB0.00 GiB69.57 GiB2.43 GiB5±8.3%
GLM-4.7-REAP-218B-A32BMoEUD-IQ1_M218B62.63 GiB6.11 GiB69.33 GiB2.67 GiB13±37%
GLM-4.5VMoEI1-Q4_K_M108B65.61 GiB3.05 GiB69.24 GiB2.76 GiB17±37%
Qwen3-235B-A22B-abliteratedMoEI1-IQ2_S235B65.40 GiB3.12 GiB69.11 GiB2.89 GiB17±37%
Devstral-2-123B-Instruct-2512IQ4_XS125B62.51 GiB5.84 GiB69.07 GiB2.93 GiB5±8.3%
grok-2MoEIQ2_XXS270B63.81 GiB4.25 GiB68.75 GiB3.25 GiB8±37%
Qwen3-VL-235B-A22B-ThinkingMoEUD-IQ1_M236B64.90 GiB3.12 GiB68.61 GiB3.39 GiB17±37%
Qwen3-VL-235B-A22B-InstructMoEUD-IQ1_M236B64.83 GiB3.12 GiB68.53 GiB3.47 GiB17±37%
gpt-oss-20b-hereticMoEIQ4_NL20.9B67.58 GiB0.41 GiB68.52 GiB3.48 GiB15±37%
GLM-4.5-Air-DerestrictedMoEQ4_1110B64.77 GiB3.05 GiB68.40 GiB3.60 GiB17±37%
GLM-4.5-AirMoEQ4_1110B64.77 GiB3.05 GiB68.40 GiB3.60 GiB17±37%
MiniMax-M2.5MoEUD-IQ1_M229B63.74 GiB4.12 GiB68.39 GiB3.61 GiB18±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ4_0124B66.12 GiB1.46 GiB68.12 GiB3.88 GiB22±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ4_1109B64.35 GiB3.19 GiB68.11 GiB3.89 GiB17±37%
HarmonicHarlequin_v5-20BQ8_033.3B32.97 GiB34.53 GiB68.10 GiB3.90 GiB5±8.3%
Qwen3.5-88BMoEI1-Q6_K87.7B67.08 GiB0.40 GiB68.05 GiB3.95 GiB24±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEI1-IQ2_S229B63.36 GiB4.12 GiB68.02 GiB3.98 GiB18±37%
step-3.5-flashIQ2_M199B59.59 GiB6.92 GiB67.09 GiB4.91 GiB5±8.3%
Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-PreservedMoEBF1635.1B66.19 GiB0.33 GiB67.07 GiB4.93 GiB27±37%
Qwen-AgentWorld-35B-A3BMoEBF1634.7B66.19 GiB0.33 GiB67.07 GiB4.93 GiB27±37%
Qwable-v1MoEBF1636.0B66.19 GiB0.33 GiB67.07 GiB4.93 GiB27±37%
Salience-1.5-ProMoEBF1636.0B66.19 GiB0.33 GiB67.07 GiB4.93 GiB27±37%
T-SearchMoEBF1636.0B66.19 GiB0.33 GiB67.07 GiB4.93 GiB27±37%
Qwen35B-Agent-R2MoEF1634.7B66.19 GiB0.33 GiB67.07 GiB4.93 GiB27±37%
Ornith-1.0-35B-Heretic-MTPMoEBF16—66.19 GiB0.33 GiB67.07 GiB4.93 GiB27±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 M2 Max run?
2071 of 2118 indexed open-weight models fit a Apple M2 Max at 32,768 context with q8_0 KV cache, the largest being Mixtral-8x22B-Instruct-v0.1 at Q3_K_L. That covers text, vision-language, image, video and speech models.
How much usable memory does a Apple M2 Max actually have?
Its nameplate is 96 GB, but about 66.96 GiB is available to a model once driver and compositor overhead is accounted for, and only 72 GB of the pool can be allocated to the GPU at all.
Is a Apple M2 Max fast for local AI?
Its memory bandwidth is 410 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.