DeepSeek R1 is an open weights model and vastly better than anything available for free including Meta’s llama. Also the model has been trained on relatively lower budget than other competitors. What this shows is that the cost of making fairly good AI models is going down exponentially and making a model that is vastly better than everyone else is going to remain a hard problem

On publicly-available IQ tests, Deepseek ties or exceeds all free AIs, but is far behind OpenAI’s models for paid users [source]

This success by DeepSeek presents interesting economic challenges. It means that the competition in foundational model providers is going to heat up and as a result no company will have an advantage to price their model usage much higher. This eats into their margins.

This basically means that the app developers who simply build a wrapper around some existing model such as Perplexity.AI would be much more poised for success and profitability than the model company itself.

Perplexity has quickly announce that DeepSeek’s US hosted version would be available for its users and more than 500 queries per day for their paid users. This means the competition in the app layer is going to heat up significantly and this bodes well for Android developers as well.