It’s either ChatGPT or Claude. These two have been our go-to models for pretty much anything. Be it, coding, content creation, or writing a professional email explaining the reason for a fake sick leave.
I mean, they are so great that they don’t give us any reason to look elsewhere. But the problem is its constraints and capabilities in the “FREE“ plan. I really hate that!!
But what if I say you can now use a better version of OpenAI and Claude Sonet 3.5 for almost free? Sounds too good to be true, right? Well! You are in for a surprise…
Introducing**Deepseek**—A company based out of China that specializes in building large-scale ultra-powerful, high-performing models and releasing them out to the public for almost free(Open-source)
When I say, “Almost Free“, I mean they do have small charges on the API which is 0.014 dollars which is way less than OpenAI or Claude
The best part is that Deepseek’s models are purely open-source. You can run it on your machine and exploit it to the fullest. The link is given below!
Link: https://github.com/deepseek-ai/DeepSeek-V3
Now, the performance:
I myself was surprised when I got to know that DeepSeek-V3, one of their models performed better than Claude Sonet 3.5 and much better than OpenAI 4o—sharing the graph below!
Well, they didn’t stop at that!
Few days back, to make it even more unbelievable, they launched DeepSeek-R1.
And again! It’s entirely open-source
They claim that DeepSeek seems to be performing on par with o1, OpenAI’s most advanced model
Explore DeepSeek Chat Website here: https://chat.deepseek.com
Okay then, why it’s so great?
DeepSeek models use the **Mixture Of Experts(MOE)**architecture. DeepSeek V3 or R1 is not a single big model. It’s a collection of multiple specialized sub-models, each designed, trained, and optimized to handle a specific subset of tasks.
The model will simply call upon the necessary sub-models based on the prompts/requests. This way, it doesn’t have to use the entire model’s capabilities, consumes less computational resources, and of course, it has faster performance.
What more do you need?…
But there’s a catch. In order to compete or perform even better than OpenAI or Claude, they need what?
More computing power and more hardware resources.
This brings me to my next point!
The Fear:
Both OpenAI and Claude have access to NVIDIA's most powerful chips A100 and H100 GPUs. So, they were able to increase the parameters to trillion and train the model for better results. They do have access because they are American companies.
But DeepSeek is a Chinese company.
So, due to U.S. export restrictions, DeepSeek has faced challenges in accessing NVIDIA's most advanced chips. In colloquial terms, they are not allowing DeepSeek to have anything from the US, because they are ______(You fill in the blanks)
I can only imagine what’s running in the minds of the people in OpenAI and Claude.
But, hypothetically. Let’s say DeepSeek somehow got access to those powerful chips of Nvidea.
If DeepSeek can outperform the two biggest AI companies with just 671 billion parameters and limited resources, how unstoppable could they be with access to cutting-edge NVIDIA chips?
That’s a Billion Dollar Question…