A few days ago, a Chinese research company called Deepseek unveiled their newest model - Deepseek R1.
R1 is a completely open source reasoning model - and one that takes a far more flexible approach than OpenAI - and beats it on both performance and price.
Before we dive into the open source parts, let's compare their hosted counterparts.
OpenAI's o1 model is their biggest model yet, with it's claim to fame being it's ability to handle complex reasoning.
For general consumers, you must pay either:
- $20 USD a month to access o1 and o1 mini on limited rate limits
- $200 USD a month to access o1 Pro Mode - o1 with more compute power for research tasks.
And for developers:
To use o1 through the OpenAI API, it will cost the developer $15 USD for 1 million input tokens, and $60 USD for 1 million output tokens.
o1 Mini is similar but a little cheaper, at $3/$12 respectively.
Deepseek R1 on the other hand, is completely free for anyone to use on https://chat.deepseek.com/ - and you can combine it with search. That is due to a REALLY cool flexibility in their model, but we'll return to that.
What about the API?
Well it's about 95% cheaper than OpenAI.
$0.55 USD for a million tokens in, and $2.19 for a million out.
So they've crushed them in price - but what about performance?
They're beating, or come very close to o1 on a LOT of metrics - and the reason why is the secret sauce in this model.
R1 is not one model, but two.
R1, in itself, is an open source reasoning layer that you can distill onto any other LLM model, no matter how big or small. That means you can greatly improve the performance and output of even the smallest models, and you can also run R1 on much less hardware. It also allows them to use features like file analysis, web searching and more - far before OpenAI and their big, bloated models can.
The DeepSeek API is also entirely OpenAI compatible, so you could execute it as a drop in replacement as simply as changing a URL and API token.
This all testifies very well against my thoughts that the future of AI is not the big super models the US wants - the future of AI is small.
People have already distilled the power of r1 with a smaller base model to run on an iPhone. And remember - this model is competing with OpenAI, the same company who just got $500 billion in funding to build infrastructure.
And we beat them with an iPhone.
This release is an incredible act of generosity in restoring power to the people and researches who want truly open AI, and will push us a lot further than OpenAI is with their for-profit visions.
You can read about the model - and run it yourself here: https://github.com/deepseek-ai/DeepSeek-R1