Apple · apple

Apple M3 Ultra

Apple M3 Ultra has 96 GB of unified memory at 819 GB/s — about 66.96 GiB usable after driver and compositor overhead. 2071 of 2118 indexed models fit at 16K 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
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
text 1781vision language 186image 2audio asr 39audio tts 21video 16embedding 26

What fits at 16K context

largest quantization that fits, per model · 2071 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Behemoth-X-123B-v2Q4_K_M123B68.19 GiB2.92 GiB71.82 GiB0.18 GiB9±8.3%
Mistral-Large-Instruct-2411Q4_K_M123B68.19 GiB2.92 GiB71.82 GiB0.18 GiB9±8.3%
MiniMax-M2.1-REAP-139B-A10BMoEI1-IQ4_XS139B69.15 GiB2.06 GiB71.74 GiB0.26 GiB33±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-IQ4_XS139B69.15 GiB2.06 GiB71.74 GiB0.26 GiB33±37%
Mistral-Medium-3.5-128BQ4_K_S128B68.01 GiB2.92 GiB71.64 GiB0.36 GiB9±8.3%
MiniMax-M2.5MoEUD-IQ2_XXS229B69.03 GiB2.06 GiB71.63 GiB0.37 GiB38±37%
MiniMax-M2.1MoEUD-IQ2_XXS229B68.98 GiB2.06 GiB71.57 GiB0.43 GiB38±37%
MiniMax-M2MoEUD-IQ2_XXS229B68.92 GiB2.06 GiB71.51 GiB0.49 GiB38±37%
c4ai-command-r-plus-08-2024Q5_K_M104B68.57 GiB2.13 GiB71.42 GiB0.58 GiB9±8.3%
OYM-Qimi-122B-A10B-K2.6MoEI1-Q4_K_M125B70.64 GiB0.20 GiB71.41 GiB0.59 GiB45±37%
Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliteratedMoEQ4_K_M123B70.63 GiB0.20 GiB71.41 GiB0.59 GiB45±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ5_K_S109B69.16 GiB1.59 GiB71.33 GiB0.67 GiB33±37%
Llama-3_3-Nemotron-Super-49B-v1_5Q8_049.9B49.36 GiB21.25 GiB71.30 GiB0.70 GiB9±8.3%
Valkyrie-49B-v2.1Q8_049.9B49.36 GiB21.25 GiB71.30 GiB0.70 GiB9±8.3%
Llama-3_3-Nemotron-Super-49B-v1Q8_049.9B49.36 GiB21.25 GiB71.30 GiB0.70 GiB9±8.3%
Step-3.7-FlashQ2_K_L201B66.82 GiB3.74 GiB71.13 GiB0.87 GiB9±8.3%
GLM-4.5-Air-DerestrictedMoEQ4_K_L110B68.88 GiB1.53 GiB70.99 GiB1.01 GiB33±37%
GLM-4.7-REAP-218B-A32BMoEUD-IQ2_XXS218B67.13 GiB3.05 GiB70.78 GiB1.22 GiB27±37%
GLM-4.5-AirMoEQ4_K_M110B68.45 GiB1.53 GiB70.56 GiB1.44 GiB33±37%
dots.llm1.instMoEIQ3_XS143B61.71 GiB8.23 GiB70.52 GiB1.48 GiB24±37%
Qwen3.5-122B-A10BMoEQ4_K_S125B69.66 GiB0.20 GiB70.44 GiB1.56 GiB45±37%
Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16MoEQ8_035.1B69.57 GiB0.17 GiB70.30 GiB1.70 GiB45±37%
Seed-OSS-36B-InstructBF1636.2B67.35 GiB2.13 GiB70.12 GiB1.88 GiB10±8.3%
Hermes-4.3-36BBF1636.2B67.35 GiB2.13 GiB70.12 GiB1.88 GiB10±8.3%
Mixtral-8x22B-Instruct-v0.1MoEQ3_K_L141B67.60 GiB1.86 GiB70.07 GiB1.93 GiB17±37%
MiMo-V2-FlashMoEKV unresolvedIQ2_XXS310B68.47 GiB1.00 GiB70.07 GiB1.93 GiB42±37%
Mixtral-8x22B-v0.1MoEQ3_K_L141B67.60 GiB1.86 GiB70.06 GiB1.94 GiB17±37%
Mixtral-8x22B-v0.1MoEQ3_K_L141B67.60 GiB1.86 GiB70.06 GiB1.94 GiB17±37%
Laguna-S-2.1MoEQ4_1118B68.96 GiB0.47 GiB70.00 GiB2.00 GiB42±37%
Devstral-2-123B-Instruct-2512Q4_K_S125B66.36 GiB2.92 GiB69.98 GiB2.02 GiB10±8.3%
XORTRON-NXTXPRTXXLI1-Q4_K_S128B66.36 GiB2.92 GiB69.98 GiB2.02 GiB10±8.3%
Qwen3.5-122B-A10B-hereticMoEI1-Q4_K_M123B69.11 GiB0.20 GiB69.89 GiB2.11 GiB46±37%
step-3.5-flashQ2_K_L199B65.26 GiB3.74 GiB69.57 GiB2.43 GiB10±8.3%
HunyuanImage-2.1Q6_K17.5B68.97 GiB0.00 GiB69.57 GiB2.43 GiB10±8.3%
Mistral-Small-4-119B-2603MoEUD-Q4_K_M119B68.70 GiB0.19 GiB69.47 GiB2.53 GiB46±37%
gpt-oss-120b-Uncensored-xCloudMoEI1-Q4_1117B68.42 GiB0.31 GiB69.27 GiB2.73 GiB45±37%
gpt-oss-120b-abliteratedMoEI1-Q4_1117B68.42 GiB0.31 GiB69.27 GiB2.73 GiB45±37%
GLM-4.6VMoEQ4_K_L108B66.89 GiB1.53 GiB68.99 GiB3.01 GiB34±37%
MiMo-V2.5MoEKV unresolvedIQ1_M311B67.01 GiB1.00 GiB68.60 GiB3.40 GiB43±37%
gpt-oss-20b-hereticMoEIQ4_NL20.9B67.58 GiB0.21 GiB68.32 GiB3.68 GiB27±37%
Step-3.5-Flash-REAP-121B-A11BI1-Q4_K_S121B63.83 GiB3.74 GiB68.14 GiB3.86 GiB10±8.3%
MiniMax-M2.7MoEUD-IQ2_M229B65.32 GiB2.06 GiB67.92 GiB4.08 GiB39±37%
Qwen3.5-88BMoEI1-Q6_K87.7B67.08 GiB0.20 GiB67.85 GiB4.15 GiB42±37%
CalmeRys-78B-Orpo-v0.1Q6_K78.0B64.27 GiB2.86 GiB67.81 GiB4.19 GiB10±8.3%
calme-2.3-rys-78bQ6_K78.0B64.27 GiB2.86 GiB67.81 GiB4.19 GiB10±8.3%
GLM-4.5VMoEI1-Q4_K_M108B65.61 GiB1.53 GiB67.71 GiB4.29 GiB34±37%
Qwen2.5-Coder-32B-InstructQ8_032.8B64.86 GiB2.13 GiB67.63 GiB4.37 GiB10±8.3%
Qwen3-235B-A22B-abliteratedMoEI1-IQ2_S235B65.40 GiB1.56 GiB67.55 GiB4.45 GiB34±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ4_0124B66.12 GiB0.73 GiB67.39 GiB4.61 GiB41±37%
Qwen3-VL-235B-A22B-ThinkingMoEUD-IQ1_M236B64.90 GiB1.56 GiB67.05 GiB4.95 GiB34±37%
Qwen3-VL-235B-A22B-InstructMoEUD-IQ1_M236B64.83 GiB1.56 GiB66.97 GiB5.03 GiB34±37%
Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-PreservedMoEBF1635.1B66.19 GiB0.17 GiB66.91 GiB5.09 GiB47±37%
Qwen-AgentWorld-35B-A3BMoEBF1634.7B66.19 GiB0.17 GiB66.91 GiB5.09 GiB47±37%
Qwable-v1MoEBF1636.0B66.19 GiB0.17 GiB66.91 GiB5.09 GiB47±37%
Salience-1.5-ProMoEBF1636.0B66.19 GiB0.17 GiB66.91 GiB5.09 GiB47±37%
T-SearchMoEBF1636.0B66.19 GiB0.17 GiB66.91 GiB5.09 GiB47±37%
Qwen35B-Agent-R2MoEF1634.7B66.19 GiB0.17 GiB66.91 GiB5.09 GiB47±37%
Ornith-1.0-35B-Heretic-MTPMoEBF16—66.19 GiB0.17 GiB66.91 GiB5.09 GiB47±37%
Ornith-1.0-35BMoEBF1634.7B66.19 GiB0.17 GiB66.91 GiB5.09 GiB47±37%
Carnice-Qwen3.6-MoE-35B-A3BMoEF1636.0B66.19 GiB0.17 GiB66.91 GiB5.09 GiB47±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.

Measured on this card

third-party benchmarks, aggregated
Workload◍MedianMiddle 50%Runs
Prompt processing1471.24 tok/s1116.66–1488.1811
Text generation64.16 tok/s52.90–92.0411
Benchmarked· n=11

Aggregated from community-submitted runs, so the spread is wide by nature — it covers different models, resolutions, step counts and settings, not one controlled configuration. Read the middle 50% rather than the median alone. These figures are reproduced with attribution from llama.cpp-discussion-10879.

Questions people ask

What AI models can a Apple M3 Ultra run?
2071 of 2118 indexed open-weight models fit a Apple M3 Ultra at 16,384 context with q8_0 KV cache, the largest being Behemoth-X-123B-v2 at Q4_K_M. That covers text, vision-language, image, video and speech models.
How much usable memory does a Apple M3 Ultra 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 M3 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.