We started building Deep Research over a year ago, driven by progress in reinforcement learning and reasoning models, initially focusing on math, science, and coding tasks.

We trained a version of GPT-3 specifically for web browsing and data analysis to enhance the capabilities of Deep Research.

The model that powers Deep Research is a fine-tuned version of the O3 model, specifically trained on challenging browsing and reasoning tasks, enhancing its online search and analysis capabilities.

Through thousands of reinforcement learning steps, DeepSeek's model learned advanced reasoning skills, including extended chains of thought and the ability to recognize and correct its own mistakes.

Deep reinforcement learning, which DeepMind is known for, allows systems to learn from scratch by trial and error, achieving superhuman performance in some tasks.

We aimed to create a useful model by training it directly on tasks that users perform in their daily lives, with online browsing being a key focus due to its broad applicability.

DeepSeek's R1 model is one of the first large models to achieve top-tier results purely through reinforcement learning, marking a significant milestone in AI development. Additionally, DeepSeek introduced a cold start phase for fine-tuning on structured reasoning examples before reinforcement learning, which eliminated language mixing issues and made outputs far more comprehensible.

Personalized education is a promising use case for Deep Research, as it can compile information on topics users want to learn about, creating tailored reports to aid in their understanding.

DeepSeek's R1 reasoning model achieves impressive performance by applying reinforcement learning specifically to think step by step through complex problems, similar to OpenAI's approach with their model, o1.