Mac Studio first test: Can a top-of-the-line Mac worth 100,000 yuan run the full version of DeepSeek?

Written by

Audrey Miles

Updated on:July-13th-2025

Since Apple developed its own M series chips, we have heard it once every year:

This is the most powerful Mac ever!

But how strong is the strongest? In the past, we could only put a few more engineering files into the Mac to show its excess performance. But this year, things have begun to change. A new generation of top-of-the-line PC testing software is no longer Mark or Bench, but DeepSeek.

The newly launched M3 Ultra is the most powerful Mac ever. It is one of the few computers that you can deploy the "full-powered version" of DeepSeek right after you buy it. Here are our actual test experiences:

Due to space limitations, the test results are not shown in the video, so we will make some supplements here:

During the test, the noise control and heat dissipation of Mac Studio were relatively good. Even when it took up 400GB of unified memory to run a large model, it did not overheat or make excessive noise.

After multiple rounds of actual testing, there is no significant difference in the processing capabilities of the Q4 quantized 671B DeepSeek large model and the 70B version on most tasks. However, the load is lower and the token throughput is faster when running the 70B model. Therefore, we recommend that you deploy the DeepSeek model locally and give priority to the tuned 70B version.

The current factory system version of Mac Studio is macOS Sequoia 15.3 (24D2059), which is not the latest macOS system and may still have certain compatibility issues.

During the testing process, we also encountered compatibility issues with Mac Studio. We installed the latest Blender software and demo, but it ran slowly. When we looked at the system usage, we found that 98% of the performance was idle.

Apple said that it will release the first day update on March 12, and it is recommended to upgrade as soon as possible.

Mac Studio is equipped with 6 Thunderbolt 5 ports with a data transmission speed of up to 120Gb/s. Before the test, we downloaded the large model file to the local hard disk and transferred it through the sugar factory's latest 80Gb/s Hummingbird cable. It only took 10 seconds to transfer the 60GB large model file.

Here's a tip for transferring files between Macs at high speed: connect two computers with a USB-C cable, completely shut down the Mac containing the files to be transferred, hold down the power button until "Loading Startup Options" appears, select the startup disk, select Utilities > Shared Disk > Start Sharing, then open Finder on another Mac and find the shared disk in the "Network" window in the sidebar to transfer files directly (this method is only available on computers with Apple chips).

If you really need to deploy the full-blooded version of DeepSeek locally, the currently recommended configuration is to purchase a Mac Studio with 512GB unified memory + 1TB SSD and an M3 Ultra chip. The official website price is 74,249 yuan, and you can add national subsidies, education discounts, 12 installments, etc. This may be the lowest threshold and most cost-effective full-blooded version of DeepSeek local deployment solution.

We do not recommend impulse purchases by ordinary consumers . From the test results, the official web version of DeepSeek is still the most stable and efficient version. There are a large number of tools on the market that can be used as a substitute (such as Tencent Yuanbao, Huawei/vivo/OPPO/nubia and other mobile phones also have built-in applications). You can also consider calling API deployment through tools such as Silicon Mobility. Even if there is a need for local deployment, or the 14B/32B/70B version is selected, for most tasks, the current 32B parameter DeepSeek/QWQ large model is quite sufficient.

For enterprises, although Mac Studio can indeed deploy large models with 600 billion parameters, it still lags behind dedicated large-model all-in-one machines in terms of memory bandwidth, model fine-tuning services, etc., and purchases require careful evaluation.

In general, we are still happy about the appearance of the 2025 Mac Studio. The appearance of this product means that professional productivity tools have reached a new level. For all creators, the substantial iteration of productivity tools and the substantial reduction in costs are always a good thing. The powerful performance of Mac Studio has greatly lowered the threshold for creation in the AI era.

You can now configure the world's most advanced LLM large model at home, which can help you complete your creation efficiently without leaving your home; you can connect up to 8 5K monitors and use open source software such as Blender to create works that are no less than those of the Hollywood film and television industry.

More and more independent teams are rapidly becoming unicorns with valuations of over 100 million US dollars. The team of midjourney, a graphic model company, has only 49 people, but its annual revenue exceeds 100 million US dollars. The team of Cursor, an AI-assisted programming software, has only 30 people, and it took only 21 months from scratch to achieve rapid profitability. The Oscar-winning animated feature film "Life of Cats" was created entirely with Blender. The ten-person team from Latvia made this small Eastern European country win this honor for the first time.

OpenAI CEO Sam Altman once predicted:

With AI, we will soon see the first AI company with a valuation of $1 billion and only one person.

For creators, now is the best time.