DeepSeek has launched R1-Lite-Preview, a reasoning model to compete with OpenAI's o1
DeepSeek, a Chinese AI lab that offshoot from quantitative hedge fund High-Flyer Capital Management and that focuses on open-source AI. DeepSeek R1-Lite-Preview is currently available on DeepSeek Chat only, and according to the announcement in X, the model performs at o1's level in the MATH and AIME benchmark tests. Unlike OpenAI, which decided to hide o1's chains of thought from users, R1-Lite-Preview's thought processes are fully transparent, allowing users to see them in real time. DeepSeek reportedly plans to open-source the model and release an API.
R1-Lite-Preview leverages the new test-time or inference compute technique, which essentially consists in providing the model with more resources (time and processing power) when it is generating outputs to user prompts, or as it is often described, giving models more time to "think" through problems before providing an answer. As often happens with new model releases, the community has been eager to test R1-Lite-Preview's capabilities and limitations, finding that much like o1, R1-Lite-Preview struggles at simple problems like tic-tac-toe. Notorious LLM hacker Pliny the Prompter quickly announced R1-Lite-Preview had been successfully jailbroken and made to generate a detailed meth recipe.