Congratulations on buying another DeepSeek all-in-one machine

Written by

Caleb Hayes

Updated on:July-15th-2025

Personally, I think that customers who do not choose the full-blooded version of 671B and insist on choosing the so-called 70B or 32B are SB customers and Party B should stay away from them. I also call on Party B to stay away from these customers.

Reasons to stay away : Even the full-blooded version may not give Party A a good experience. After the non-full-blooded version is expected to be launched, Party A will definitely have a bad experience. Party A is definitely not complaining about his own wrong choice, but Party B's system is not good. Party B is wronged and it ruins Party B's own reputation. But then again, Party B deserves the bad luck, because you accepted it.

《Top 10 DeepSeek all-in-one machine requirements from 100+ customers》

What is the threshold for DeepSeek all-in-one machine?

DeepSeek all-in-one organization promotes an industry scale of 500 billion in 2 years. My God, this industry is just like real estate agency . The entry threshold is very low, but it is really difficult to do it well , especially under limited computing power conditions, to achieve R1 IQ without reducing or reducing the official website at the minimum.

Many people say that it is enough for customers to choose 32B and 70B. I want to tell you that today's customers have already experienced the high-IQ version of R1 on the public network . If you deploy a 70B for them and let them use it on the intranet, can they really not curse the other party in their hearts? Do they really think it is easy to use? !

It is easy to go from frugality to luxury, but difficult to go from luxury to frugality. Today's netizens have experienced the highly intelligent R1. The IT era of relying on deception to get orders is over. In the AI 1.0 era, everyone could rely on deceiving customers to try new things and let them pay for trial and error. But now it is the A2.0 era. No one's money comes from the wind. If you deceive customers to pay, but the hardware and software systems cannot be used, won't you feel guilty?

Why do domestic AI chip manufacturers promote 32B and 70B

1. AI chips have poor computing power and can only be deployed in 32B and 70B. The hardware cost of deploying 671B is too high and customers simply cannot afford it, so they can only use the small parameter version.

For example, some domestic AI chip manufacturers need 200,000 RMB to deploy a 32B machine with 4 cards, and 300,000 RMB to deploy a 70B machine with 8 cards. Instead of focusing on improving the cost-effectiveness of AI chip iteration, they focus on how to take advantage of hot spots and harvest customers. How can they make good AI chips in this way?

2. To reduce inventory pressure , Deepseek reduced the demand for large model inference, allowing some small computing chips to be shipped. Many AI manufacturers began to think of ways to ship, such as using A chips + large video memory to make some new AI cards, and letting channels promote 32B and 70B versions, so-called letting customers try out new products. Such customers will definitely end up with a mess. How can they use a stupid all-in-one machine?! Is this a taste of new products or...

I won’t talk about the other less harmonious ones.

Why do software vendors promote 32B and 70B?

Now every software manufacturer is making DeepSeek large model all-in-one machine, and every AI hardware manufacturer is also making DeepSeek large model all-in-one machine. Is this a good thing or a bad thing? !

The 671B on the official website is particularly suitable for cluster operation. If you want to run a stand-alone version, there is actually a lot of optimization work to do, otherwise the concurrency will not be able to be increased. Even for quantitative deployment, even if the code is not changed and only the configuration is explored, many teams that have not trained large models cannot do it. They can only find some information on the Internet and try it themselves. So since it is so difficult, just choose to promote 32B and 70B to customers, at least it is simple.

There are also software manufacturers who have good customer relationships based on their previous customer resources. After all, they have transaction records. Now they are pushing it hard, taking advantage of the popularity, giving customers a gimmick first, and directly buying a minimum version, tricking customers into going online first, thus occupying the trap.

What’s more, I have two partners who directly sent 3 machines to 3 customers. The customers went online first, and they sent a message to push the machine online first, and then the operation and maintenance fees were settled later. I really am...

How to choose an all-in-one machine?

Let me just say one thing. Practice is the only criterion for testing truth . Testing IQ is very simple. Ask the same question to the official website and the all-in-one machine and see how big the difference in replies is. That means your IQ has dropped a lot. Congratulations, you have bought another stupid DeepSeek all-in-one machine.

Compare these 5 questions, or expand them to 10 questions and you will immediately judge your IQ.

Question 1: Which is bigger, 7.11 or 7.9?

Question 2: There are three liars A/B/C. One of them will tell the truth if and only if the other two are lying at the same time. Please establish a nonlinear system of equations to describe their relationship.

Question 3: Assuming you are an AI consultant for a country’s central bank, please design a monetary policy that allows the legal circulation of cryptocurrencies while maintaining inflation targeting and preventing the impact of quantum computers on traditional encryption systems.

Question 4: Combine quantum biology, computational neuroscience, and phenomenology to explain the mechanism of human consciousness and propose an experimental plan to verify your theory, which must include falsifiability criteria.

Question 5: If you are asked to decipher Linear A, please design a multimodal neural network architecture that integrates archaeological background knowledge, pottery pattern analysis, and syllable statistical features to provide a deciphering roadmap.