YouTuber DaveLee demonstrated how Apple's new MacStudio equipped with an M3Ultra chip can efficiently run a huge version of the DeepSeekR1AI model locally, provided the user equips the machine with a maximum of 512GB of memory.
According to Lee's tests, the 671 billion-parameter AI model can be executed directly on Apple's high-end workstations, but it requires a large amount of memory resources, consumes 404GB of storage space, and requires manual allocation of 448GB of virtual memory through terminal commands.
M3Ultra's unified memory architecture is key to achieving this performance, allowing the system to efficiently process the 4-bit quantized version of DeepSeekR1. Quantization reduces the accuracy slightly, but it maintains all parameters and provides about 17-18 marks per second, which is sufficient for many practical applications.
Most impressively, MacStudio achieves this performance while consuming less than 200 watts of power. Achieving comparable performance on traditional PC hardware would require multiple GPUs and consume approximately ten times the power.
The ability to run such advanced AI models locally provides privacy benefits for sensitive applications such as healthcare data analysis, where sending information to cloud services raises security concerns.
That kind of performance doesn't come cheap, though—Mac Studio starts at about $10,000 with M3 Ultra and 512GB of memory. If the performance of M3UltraMacStudio is maximized, the price of the Apple M3Ultra chip equipped with 16TB SSD storage and 32-core CPU, 80-core GPU and 32-core neural engine will be as high as $14,099.
Apple says M3 Ultra is the fastest Mac chip it's ever released, thanks to the company's strategy of fusing two M3 Max chips together using "UltraFusion" technology. This makes the chip twice the specs of the M3Max.