one year on
DeepSeek releases R1-Lite-Preview with o1-preview-level performance and full chain-of-thought transparency
The Chinese AI lab matches OpenAI's o1-preview on math benchmarks while showing its every step in real time, and promises open-source release.
DeepSeek today released R1-Lite-Preview, a reasoning model that matches OpenAI’s o1-preview on AIME and MATH benchmarks while displaying its complete chain-of-thought reasoning in real time. The model is available at chat.deepseek.com, and the company says open-source weights and an API are coming soon.
The model shows inference scaling laws: performance on AIME improves steadily as the model is allowed to reason longer. This transparency differs from OpenAI’s o1-preview. DeepSeek’s approach lets users follow every logical step, making the model potentially useful for education, problem-solving, and research.
DeepSeek’s open-source promise could widen access to advanced reasoning capabilities.
The record
Announced the model on X, highlighting o1-preview-level performance on AIME and MATH benchmarks and promising open-source models and API coming soon.
One year later — open only if you can handle spoilers
Two months later, DeepSeek's full R1 release triggered a major market sell-off as investors realized the gap between US and Chinese AI capabilities had nearly closed. The warning shot from R1-Lite-Preview was largely ignored by Western analysts focused on OpenAI's lead.