In January 2025, Perplexity announced the integration of the DeepSeek R1 model into its platform, potentially bringing revolutionary change to AI-based searches. The Chinese-developed model, which runs exclusively on American and European servers, is not only more cost-effective than its competitors but also outperforms them in terms of performance whilst guaranteeing data security for Western users.
The DeepSeek R1 model has 671 billion parameters and costs just $0.14 per token, compared to the $7.50 price of OpenAI's o1 model. Its performance is outstanding: it achieved a 97.3% result on MATH-500 tests, whilst o1 achieved only 78.5%. Furthermore, thanks to Cerebras AI infrastructure, R1 can process 1,500 tokens per second, representing a 57-fold speed compared to traditional GPU-based systems. Aravind Srinivas, CEO of Perplexity, emphasised that no user data reaches China as the service operates exclusively on Western servers.
On 21st February 2025, Perplexity further increased the availability of the R1 model by open-sourcing R1 1776, a censorship-free version. The new model is available through the Hugging Face platform and the Perplexity Sonar API. According to the company, the R1 1776 model has been optimised to provide impartial and factual information whilst preserving its excellent mathematical and logical capabilities. Perplexity's subscription model has also undergone significant changes. Pro subscribers can perform 500 searches per day using DeepSeek R1, while free users can access three Pro-level queries daily. This step represents an important advancement in ensuring the widespread availability of advanced AI technologies.
Sources:
1.
Learn about DeepSeek-R1, an open-source advanced reasoning language model integrated into Perplexity's Pro Search, offering enhanced logical inference and decision-making capabilities.
2.
Radhika Rajkumar
3.
Daniel Ferrera
