Skip to content

Mac Studio M4 Ultra 512GB

Apple · M4 Ultra · 512GB Unified Memory · Can run 114 models

BuyApple
ManufacturerApple
Unified Mem512 GB
ChipM4 Ultra
CPU Cores32
GPU Cores80
Neural Engine32
Bandwidth819 GB/s
MSRP$9,999
ReleasedMar 12, 2025

AI Notes

The Mac Studio M4 Ultra 512GB is the ultimate local AI machine. With a staggering 512GB unified memory and 819 GB/s bandwidth, it can run the largest open-source models including 405B parameter models without quantization. It rivals multi-GPU server setups at a fraction of the complexity.

Compatible Models

ModelParametersBest QuantVRAM UsedFitEst. Speed
Qwen 3 0.6B600MQ4_K_M2.5 GBRuns~328 tok/s
Qwen 3.5 0.8B800MQ4_K_M1.5 GBRuns~546 tok/s
Gemma 3 1B1BQ8_02 GBRuns~410 tok/s
Llama 3.2 1B1BQ8_03 GBRuns~273 tok/s
DeepSeek R1 1.5B1.5BQ8_03 GBRuns~273 tok/s
SmolLM2 1.7B1.7BQ8_02.7 GBRuns~303 tok/s
Gemma 2 2B2BQ8_04 GBRuns~205 tok/s
Gemma 3n E2B2BQ4_K_M3.3 GBRuns~248 tok/s
Gemma 4 E2B2BQ4_K_M4 GBRuns~205 tok/s
Qwen 3.5 2B2BQ4_K_M3 GBRuns~273 tok/s
Llama 3.2 3B3BQ8_05 GBRuns~164 tok/s
StarCoder2 3B3BQ4_K_M3.5 GBRuns~234 tok/s
Phi-3 Mini 3.8B3.8BQ8_05.8 GBRuns~141 tok/s
Phi-4 Mini 3.8B3.8BQ4_K_M4.5 GBRuns~182 tok/s
Gemma 3 4B4BQ4_K_M5 GBRuns~164 tok/s
Gemma 3n E4B4BQ4_K_M4.5 GBRuns~182 tok/s
Gemma 4 E4B4BQ4_K_M6 GBRuns~137 tok/s
Qwen 3 4B4BQ4_K_M4.5 GBRuns~182 tok/s
Qwen 3.5 4B4BQ4_K_M4.5 GBRuns~182 tok/s
Yi 1.5 6B6BQ4_K_M5 GBRuns~164 tok/s
Codestral Mamba 7B7BQ4_K_M6.9 GBRuns~119 tok/s
DeepSeek R1 7B7BQ8_09 GBRuns~91 tok/s
Falcon 3 7B7BQ4_K_M6.8 GBRuns~120 tok/s
InternLM 2.5 7B7BQ4_K_M5.5 GBRuns~149 tok/s
Mistral 7B7BQ8_09 GBRuns~91 tok/s
OpenChat 3.5 7B7BQ4_K_M6.9 GBRuns~119 tok/s
Qwen 2.5 7B7BQ8_09 GBRuns~91 tok/s
Qwen 2.5 Coder 7B7BQ8_09 GBRuns~91 tok/s
Qwen 2.5 VL 7B7BQ4_K_M7 GBRuns~117 tok/s
StarCoder2 7B7BQ4_K_M5.5 GBRuns~149 tok/s
WizardLM 2 7B7BQ4_K_M6.9 GBRuns~119 tok/s
Aya Expanse 8B8BQ4_K_M6.5 GBRuns~126 tok/s
Cogito 8B8BQ4_K_M7.5 GBRuns~109 tok/s
DeepSeek R1 8B8BQ4_K_M7.5 GBRuns~109 tok/s
Dolphin 3 8B8BQ4_K_M6 GBRuns~137 tok/s
Granite 3.3 8B8BQ8_010 GBRuns~82 tok/s
Llama 3.1 8B8BQ8_010 GBRuns~82 tok/s
Nemotron 3 Nano 8B8BQ4_K_M7.5 GBRuns~109 tok/s
Nous Hermes 2 8B8BQ4_K_M6 GBRuns~137 tok/s
Qwen 3 8B8BQ4_K_M7.5 GBRuns~109 tok/s
Gemma 2 9B9BQ8_011 GBRuns~74 tok/s
Qwen 3.5 9B9BQ4_K_M7.5 GBRuns~109 tok/s
Yi 1.5 9B9BQ4_K_M6.5 GBRuns~126 tok/s
Yi Coder 9B9BQ4_K_M8 GBRuns~102 tok/s
Falcon 3 10B10BQ4_K_M8.5 GBRuns~96 tok/s
Llama 3.2 Vision 11B11BQ4_K_M8.5 GBRuns~96 tok/s
Gemma 3 12B12BQ4_K_M10.5 GBRuns~78 tok/s
Mistral Nemo 12B12BQ4_K_M9.5 GBRuns~86 tok/s
DeepSeek R1 14B14BQ4_K_M9.9 GBRuns~83 tok/s
Phi-4 14B14BQ4_K_M9.9 GBRuns~83 tok/s
Phi-4 Reasoning 14B14BQ4_K_M11 GBRuns~74 tok/s
Qwen 2.5 14B14BQ4_K_M9.9 GBRuns~83 tok/s
Qwen 2.5 Coder 14B14BQ4_K_M12 GBRuns~68 tok/s
Qwen 3 14B14BQ4_K_M12 GBRuns~68 tok/s
StarCoder2 15B15BQ8_017 GBRuns~48 tok/s
InternLM 2.5 20B20BQ4_K_M12 GBRuns~68 tok/s
gpt-oss 20B21BMXFP415 GBRuns~55 tok/s
Codestral 22B22BQ4_K_M14.7 GBRuns~56 tok/s
Devstral 24B24BQ4_K_M17 GBRuns~48 tok/s
Magistral Small 24B24BQ4_K_M17 GBRuns~48 tok/s
Mistral Small 3.1 24B24BQ4_K_M18 GBRuns~46 tok/s
Gemma 4 26B26BQ4_K_M20 GBRuns~41 tok/s
Gemma 2 27B27BQ4_K_M17.7 GBRuns~46 tok/s
Gemma 3 27B27BQ4_K_M20 GBRuns~41 tok/s
Qwen 3.5 27B27BQ4_K_M19 GBRuns~43 tok/s
Qwen 3.6 27B27BQ4_K_M20 GBRuns~41 tok/s
Qwen 3 30B-A3B (MoE)30BQ4_K_M22 GBRuns~37 tok/s
Gemma 4 31B31BQ4_K_M22 GBRuns~37 tok/s
Aya Expanse 32B32BQ4_K_M22 GBRuns~37 tok/s
Cogito 32B32BQ4_K_M21.5 GBRuns~38 tok/s
DeepSeek R1 32B32BQ4_K_M20.7 GBRuns~40 tok/s
Qwen 2.5 32B32BQ4_K_M20.7 GBRuns~40 tok/s
Qwen 2.5 Coder 32B32BQ4_K_M23 GBRuns~36 tok/s
Qwen 3 32B32BQ4_K_M23 GBRuns~36 tok/s
QwQ 32B32BQ4_K_M21.5 GBRuns~38 tok/s
Laguna XS 2.133BQ4_K_M22 GBRuns~37 tok/s
WizardCoder 33B33BQ4_K_M22 GBRuns~37 tok/s
Nous Hermes 2 34B34BQ4_K_M19 GBRuns~43 tok/s
Yi 1.5 34B34BQ4_K_M21 GBRuns~39 tok/s
Command R 35B35BQ4_K_M22.5 GBRuns~36 tok/s
Qwen 3.5 35B A3B35BQ4_K_M12 GBRuns~68 tok/s
Qwen 3.6 35B-A3B35BQ4_K_M27 GBRuns~30 tok/s
Dolphin Mixtral 8x7B47BQ4_K_M26 GBRuns~32 tok/s
Mixtral 8x7B47BQ4_K_M29.7 GBRuns~28 tok/s
Cogito 70B70BQ4_K_M43 GBRuns~19 tok/s
DeepSeek R1 70B70BQ4_K_M43.5 GBRuns~19 tok/s
Llama 3.1 70B70BQ4_K_M43.5 GBRuns~19 tok/s
Llama 3.3 70B70BQ4_K_M43.5 GBRuns~19 tok/s
Qwen 2.5 72B72BQ4_K_M44.7 GBRuns~18 tok/s
Qwen 2.5 VL 72B72BQ4_K_M41 GBRuns~20 tok/s
Llama 3.2 Vision 90B90BQ4_K_M50 GBRuns~16 tok/s
Command R+ 104B104BQ4_K_M57 GBRuns~14 tok/s
Llama 4 Scout (109B/17B active)109BQ4_K_M72 GBRuns~11 tok/s
Command A 111B111BQ4_K_M61 GBRuns~13 tok/s
gpt-oss 120B117BMXFP470 GBRuns~12 tok/s
Qwen 3.5 122B122BQ4_K_M85 GBRuns~10 tok/s
Devstral 2 123B123BQ4_K_M67 GBRuns~12 tok/s
Mistral Large 2 123B123BQ4_K_M67 GBRuns~12 tok/s
Mixtral 8x22B141BQ4_K_M86 GBRuns~10 tok/s
Qwen 3 235B-A22B235BQ4_K_M138 GBRuns~6 tok/s
Nemotron Ultra 253B253BQ4_K_M155 GBRuns~5 tok/s
Llama 4 Maverick400BQ4_K_M228 GBRuns~4 tok/s
Llama 3.1 405B405BQ4_K_M244.5 GBRuns~3 tok/s
MiniMax M3428BQ2_K163 GBRuns~5 tok/s
DeepSeek R1 671B671BQ4_K_M362 GBRuns~2 tok/s
DeepSeek V3671BQ4_K_M362 GBRuns~2 tok/s
DeepSeek V3-0324671BQ4_K_M362 GBRuns~2 tok/s
DeepSeek V3.2671BQ4_K_M420 GBRuns~2 tok/s
GLM-5744BQ2_K300 GBRuns~3 tok/s
GLM-5.2753BQ2_K274 GBRuns~3 tok/s
GLM-5.1754BQ2_K305 GBRuns~3 tok/s
Kimi K2.61TQ2_K360 GBRuns~2 tok/s
Kimi K2.7 Code1TQ2_K359 GBRuns~2 tok/s
Kimi K2.51TQ2_K390 GBRuns~2 tok/s