Chinese AI Lab DeepSeek Challenges OpenAI with New Reasoning Model

DeepSeek has just dropped a preview of its new reasoning model, DeepSeek-R1-Lite-Preview, which the company claims is competitive with OpenAI's o1 model. It does so while offering users something rare: a window into the model's thinking. DeepSeek actually shows the model's exact step-by-step reasoning, something that not even OpenAI provides.

Details: The company hasn't yet shared a model card, detailed benchmarks, or in-depth training architecture for the R1-Lite-Preview. However, you can try the model via their web-based chatbot, DeepSeek Chat and assess it for yourself. Note, you are limited to 50 messages per day.

The big picture: This model arrives just two months after OpenAI debuted its o1-preview "reasoning" model. Unlike popular models like Claude 3.5 or Llama 3, "reasoning models" apply inference time scaling laws to solve more complex problems and provide more accurate answers. DeepSeek has managed to successfully replicate this capability in R1-Lite-Preview.

Yes, but: Like all consumer-facing models out of China, the model output is heavily censored and it declines to answer questions about sensitive political topics like Chinese leadership or Taiwan. However, it has already been jailbroken.

Looking ahead: DeepSeek plans to release open-source versions of its R1 models and the associated APIs soon, in line with their legacy of supporting the open-source AI community.