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.

The reasoning models being developed by DeepSeek are designed to utilize more compute, emphasizing the growing importance of inference in complex AI tasks.

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.

The shift from merely scaling LLMs to incorporating reasoning techniques has opened new avenues for AI development, allowing models to perform better by thinking longer and more deeply.

Reasoning is expected to be a core capability of future models, and the two paradigms of unsupervised pre-training and specialized reasoning complement each other well.

The reasoning capabilities of AI models are improving, and we encourage developers to set ambitious goals for what they can achieve with these models.

The DeepSeek paper detailed that they used reasoning-heavy but standard post-training techniques after large-scale reasoning training.

The reasoning model of DeepSeek allows for deeper answers by thinking through the steps of a problem, which was not released with previous models like OpenAI's.

The ability to allow AI models to reason for extended periods, rather than just providing quick answers, can lead to significantly better outcomes for complex questions.