Last night, the domestic large model platform DeepSeek released a preview version of the new inference model DeepSeek-R1-Lite.

The biggest feature of this model is deep thinking chain reasoning, especially in mathematics, code and various complex reasoning tasks. It can generate tens of thousands of words of reasoning process, allowing users to deeply understand the content generated by the model. The whole process.

For example, even models such as GPT-4o got it wrong that 9.11 is a bigger "dilemma" than 9.9. R1 can easily solve it through ultra-long thinking chain reasoning.

It is worth mentioning that the test data of R1 in the American Mathematics Invitational Competition AIME 2024, MATH and Codeforces were 52.5, 91.6 and 1450 respectively, defeating the o1 preview version of OpenAI, and the open source model and API will also be very popular. Release soon.

After the release of R1, it received praise from a large number of foreign netizens. Some netizens said that with the release of DeepSeek R1, OpenAI will have a strong opponent, forcing them to release the full-blooded version of o1 as soon as possible.

Amazing job! To surpass o1-preview is a huge achievement!

This is incredible! If you just use this inference model for inference tasks and use traditional language models for other things, then 50 messages a day is really enough for the average person. Well done and congratulations from Brazil!

I just tested the Deep Thinking model posted by @deepseek_ai with a highly complex research question. I was blown away by its thinking and reasoning process! In my opinion, this reaches an advanced PhD level, and in some cases the reasoning is far better than o1-preview! I was in awe.

Very good! Looking forward to your API.

Oh my God, China did it! DeepSeek has just released DeepSeek-R1-Lite-Preview. Their inference model performs as well as o1-preview*, or even better.

I tested it with some questions that only o1-preview could answer, and it worked perfectly. And it will be open sourced soon. If it happens, it will have a huge impact on the entire AI industry.

It’s really great to be able to see DeepSeek’s thinking and reasoning process.

I tried this model and it still seems to be inferior to o1-preview in terms of coding for certain tasks. But I think it's more mathematically capable. The overall performance is about the same, I really hope OpenAI releases the o1-full version now.

Real-time and transparent thought process is very important! We get to see its thought process and it's amazing.

Can there be anotherIt's always great to have one brain working together. here you go!

Great!

I was shocked. Visible thought chains are a major breakthrough for open AI research. Congratulations!

That’s crazy. When will the API be opened?

Some netizens also posted a test video of R1: the R1 model open sourced by @deepseek_ai easily 'thought' for more than 100 seconds and generated more than 7500 consecutive tokens!

It’s time to take the open source model seriously. DeepSeek just changed the game with its new model, the R1-lite. By scaling the calculations on test like o1, and 'thinking' for longer (around 5 minutes when I tried it), it achieved a state-of-the-art 91.6% on the MATH benchmark! Try it if you feel like it!

Currently, DeepSeek has not opened R1 papers, but it can be used online for free, providing 50 deep thinking reasonings every day. As the netizen above said, as long as you are not specialized in scientific research or programming development, this is enough.

The "AIGC Open Community" has experienced it and the reasoning process is indeed very strong and transparent. Let’s ask a very classic question that has caused headaches for countless big models - which one is bigger, 9.11 or 9.9.

I tried GPT-4o mini first, and still gave the wrong answer that 9.11 is greater than 9.9. This happened no matter how many times I asked.

Try R1 again. Without opening the super thinking chain, the answer is correct, 9.9 is bigger.

Try to turn on deep thinking. R1 will show all the thinking and continuous reflection process. It is very long, and the final result is still 9.9.

Currently, R1 provides 50 deep thinking chain reasonings for free every day. Interested friends can try it.

Free trial: https://chat.deepseek.com/a/chat/