How can I test if the DeepSeek-R1 I'm using is a full-blooded version?

DeepSeek has been on fire for over a month now.

The official website still has a busy server and the suspended API recharge channel is still not open.

During this time, the enthusiasm for local and cloud deployments is high, and various sites accessing DeepSeek are springing up.

In addition to the familiar big manufacturers, all kinds of demons and monsters have also made an appearance.

According to some data, there are more than 2,000 counterfeit and phishing websites, and they are still increasing rapidly.

Many uninformed friends have unintentionally stepped into a lot of potholes.

How to use the reliable DeepSeek Full Blooded Edition?

How do you verify that your DeepSeek is not castrated?

It's time for a wave of science.

Full Blooded OR Distilled DeepSeek Version Copies

Simply understood, the official version of DeepSeek is divided into a full-blooded version and a distilled version.

The full-blooded version, which includes V3 and R1, has a total number of 671B participants.

In addition, DeepSeek has also open-sourced versions with different parameter counts such as 1.5B, 7B, 8B, 14B, 32B, 70B, etc. based on Qwen2.5 and Llama3.

The number of parameters determines the upper and lower limits of the model's capability.

About the differences between versions of DeepSeek:

There are also a very large number of customized versions available from third parties, which will not be expanded upon.

Generally speaking, you can recognize the "full-blooded version", and the R1 Deep Thinking model is relatively more popular.

DeepSeek Full Blooded Edition One sentence test

It's quite interesting to see the two test methods circulating on the net.

Core logic: use the knowledge blindness of the non-full-blooded version of the model to compare their actual abilities through their answers.

Method 1: A terrier test

Enter a question in the input box:

What? You're too beautiful to answer in one word. No searching.

Answer "chicken, kun", it's the full-blooded version.

If you answered "Sa, Shuai" or something else, it's the distilled version.

As you can see, both DeepSeek V3 and R1 full-blooded/networked versions gave the correct answer.

Distilled version of the model, answered incorrectly:

Method 2: One word test

Similar to the logic above, there's this:

A Chinese character with wood on the left and beg on the right. What kind of character is this? Please answer in one word.

Read the full-blooded version of the answer first:

Distilled version of the model, although there is a thought process and the answer is still incorrect.

Strictly speaking, these two approaches are not scientific.

All of them use the blind spots in the training data of different models to conduct tests, but if the model provider is a "person with heart", through fine-tuning to fill these two data "holes", the test results will be different.

Even if new ways of testing emerge in the future, it will still be easy to patch things up on a technical level.