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 4K context with f16 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 4K context

largest quantization that fits, per model · 2071 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
dots.llm1.instMoEIQ3_M143B67.49 GiB3.88 GiB71.95 GiB0.05 GiB17±37%
Apertus-70B-Instruct-2509Q8_070.6B69.87 GiB1.25 GiB71.85 GiB0.15 GiB5±8.3%
Devstral-2-123B-Instruct-2512Q4_K_M125B69.75 GiB1.38 GiB71.84 GiB0.16 GiB5±8.3%
Mistral-Medium-3.5-128BI1-Q4_K_M128B69.75 GiB1.38 GiB71.84 GiB0.16 GiB5±8.3%
XORTRON-NXTXPRTXXLI1-Q4_K_M128B69.75 GiB1.38 GiB71.84 GiB0.16 GiB5±8.3%
GLM-4.5VMoEI1-Q5_K_S108B70.53 GiB0.72 GiB71.82 GiB0.18 GiB20±37%
Meta-Llama-3-70B-InstructQ8_070.6B69.83 GiB1.25 GiB71.76 GiB0.24 GiB5±8.3%
calme-2.4-llama3-70bQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
calme-2.2-llama3-70bQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
L3.3-Electra-R1-70bQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
L3.3-70B-Magnum-v4-SEQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
Hermes-4-70BQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
Llama-3.3_70_b_uncensored_continuedQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
grok-oss-Revenant-70BQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
Llama-3.1-70BQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
Llama-3.3-70B-InstructQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
Llama-3.1-Nemotron-70B-Instruct-HFQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
Llama-3.3-70B-Instruct-abliteratedQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
Hermes-4-70B-hereticQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
Anubis-70B-v1.2Q8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
L3.3-70B-Euryale-v2.3Q8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
Rombos-LLM-70b-Llama-3.3Q8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
llama-3-firefunction-v2Q8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
DeepSeek-R1-Distill-Llama-70B-hereticQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
DeepSeek-R1-Distill-Llama-70BQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
DeepSeek-R1-Distill-Llama-70B-abliteratedQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
Legion-V2.1-LLaMa-70BQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
Tess-R1-Limerick-Llama-3.1-70BQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
Assistant_Pepe_70BQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
SEMIKONG-70BQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
functionary-medium-v3.2KV unresolvedQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
Llama-3.1-WhiteRabbitNeo-2-70BQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
Athene-70BQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
Infinity-Instruct-7M-Gen-Llama3_1-70BQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
Hermes-3-Llama-3.1-70BQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
L3.3-70B-Magnum-DiamondQ8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
Meta-Llama-3-70B-Instruct-abliterated-v3.5Q8_070.6B69.83 GiB1.25 GiB71.75 GiB0.25 GiB5±8.3%
Step-3.7-FlashUD-IQ3_XXS201B68.54 GiB2.53 GiB71.65 GiB0.35 GiB5±8.3%
OYM-Qimi-122B-A10B-K2.6MoEI1-Q4_K_M125B70.64 GiB0.09 GiB71.31 GiB0.69 GiB26±37%
Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliteratedMoEQ4_K_M123B70.63 GiB0.09 GiB71.31 GiB0.69 GiB26±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEI1-IQ2_M229B69.70 GiB0.97 GiB71.20 GiB0.80 GiB24±37%
MiniMax-M2.1MoEI1-IQ2_M229B69.70 GiB0.97 GiB71.20 GiB0.80 GiB24±37%
MiniMax-M2.5MoEI1-IQ2_M229B69.70 GiB0.97 GiB71.20 GiB0.80 GiB24±37%
GLM-4.6-REAP-268B-A32BMoEUD-IQ1_S269B69.09 GiB1.44 GiB71.12 GiB0.88 GiB18±37%
Step-3.5-Flash-REAP-121B-A11BI1-Q4_K_M121B67.82 GiB2.53 GiB70.93 GiB1.07 GiB5±8.3%
MiniMax-M2.1-REAP-139B-A10BMoEI1-IQ4_XS139B69.15 GiB0.97 GiB70.65 GiB1.35 GiB21±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-IQ4_XS139B69.15 GiB0.97 GiB70.65 GiB1.35 GiB21±37%
Behemoth-X-123B-v2Q4_K_L123B68.47 GiB1.38 GiB70.55 GiB1.45 GiB5±8.3%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ5_K_S109B69.16 GiB0.75 GiB70.48 GiB1.52 GiB20±37%
MiniMax-M2MoEUD-IQ2_XXS229B68.92 GiB0.97 GiB70.42 GiB1.58 GiB24±37%
Qwen3.5-122B-A10BMoEQ4_K_S125B69.66 GiB0.09 GiB70.34 GiB1.66 GiB27±37%
c4ai-command-r-plus-08-2024Q5_K_M104B68.57 GiB1.00 GiB70.30 GiB1.70 GiB5±8.3%
Mistral-Large-Instruct-2411Q4_K_M123B68.19 GiB1.38 GiB70.27 GiB1.73 GiB5±8.3%
Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16MoEQ8_035.1B69.57 GiB0.08 GiB70.21 GiB1.79 GiB27±37%
GLM-4.5-Air-DerestrictedMoEQ4_K_L110B68.88 GiB0.72 GiB70.18 GiB1.82 GiB20±37%
Laguna-S-2.1MoEQ4_1118B68.96 GiB0.33 GiB69.86 GiB2.14 GiB24±37%
Qwen3.5-122B-A10B-hereticMoEI1-Q4_K_M123B69.11 GiB0.09 GiB69.78 GiB2.22 GiB27±37%
GLM-4.5-AirMoEQ4_K_M110B68.45 GiB0.72 GiB69.75 GiB2.25 GiB20±37%
HunyuanImage-2.1Q6_K17.5B68.97 GiB0.00 GiB69.57 GiB2.43 GiB5±8.3%
MiMo-V2-FlashMoEKV unresolvedIQ2_XXS310B68.47 GiB0.47 GiB69.54 GiB2.46 GiB26±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 4,096 context with f16 KV cache, the largest being dots.llm1.inst at IQ3_M. 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.