DeepSeek-R1 debuted in January 2025 in Hangzhou by an AI company. It is open source, MIT license, so anyone can look and customize it.
The monthly active users of this AI based search engine has grown globally to approximately 125 million by Q2 2025. That’s growth!
DeepSeek website traffic has also grown significantly millions of daily visitors in mid-2024 to tens of thousands daily visitors in early 2025.
The secret sauce is their Mixture of Experts system design and integrated architecture. It reasons, codes, and solves math efficiently using considerably less compute than other competitive models.
What is the training cost? The training cost is approximately US$6 million quite a bit lower compared to similar models that cost $100 million or more. Meta AI is also an option if I talk about its competitors.
The API usage is also very, very inexpensive costs approximately $0.55 per million input tokens and $2.19 per million output tokens.
Upon Marketplace release and availability on app stores, there were several stocks that had a failing day including Nvidia which dropped 17% almost $1 trillion in tech valuations all wiped out in a single day of trading.
On Reddit, one of the quotes from other users is that it “thinks too much,” and that is a good thing – it continues to peel back layers when solving “hard” problems.
Super affordable
Open source like vibe lets developers play with the app
Decent logic, code, math
Skyrocketed up app charts worldwide
Hallucinations are still there, like other AI tools
Safety challenges 1 study found ~12% unsafe outputs, versus ~1.2%; with some other model
Censorship exists within responses through policy compliance
DeepSeek-R1 is basically the underdog hero of AI search. It’s small, packed with power, and does not require much, with reasonable reasoning, and coding abilities.
It does have its issues: hallucinations, safety, censorship: those are real. But its efficiency and open source nature are a difference maker for anyone who wants smart AI for great pricing.