DeepSeek, a subsidiary of China's private equity giant Huanfang Quantitative, recently released its latest inference-focused large language model, R1-Lite-Preview. This model is currently available to the public exclusively through the DeepSeek Chat web chatbot platform.
DeepSeek is known for its innovative contributions to the open-source AI ecosystem, and this new release aims to provide the public with high-level reasoning capabilities while maintaining a commitment to accessibility and transparency. Although R1-Lite-Preview is currently only available in the chat application, it has garnered widespread attention for its performance, which is comparable to or even exceeds that of OpenAI's recently released o1-preview model.
R1-Lite-Preview employs "chain-of-thought" reasoning, demonstrating the different thought processes it undergoes when responding to user queries. While some reasoning chains may appear nonsensical or incorrect to humans, overall, R1-Lite-Preview's answers are very accurate and can even solve "trap" questions that challenge traditional powerful AI models like GPT-4o and the Claude series, such as "How many letter R's are in the word 'strawberry'?" and "Which is larger, 9.11 or 9.9?"
According to DeepSeek, the model excels in tasks requiring logical reasoning, mathematical thinking, and real-time problem-solving. Its performance surpasses that of OpenAI's o1-preview in established benchmark tests like the AIME (American Invitational Mathematics Examination) and MATH.
Additionally, DeepSeek has released extended data for the model, showing a steady increase in accuracy when given more time or "thinking tokens" to solve problems. The charts emphasize that as the depth of thought increases, the model's scores on benchmarks like AIME improve.
Currently, the release of R1-Lite-Preview performs excellently on key benchmarks, handling a range of tasks from complex mathematics to logical scenarios, scoring comparably to top reasoning models like GPQA and Codeforces. The model's transparent reasoning process allows users to observe its logical steps in real time, enhancing the system's accountability and credibility.
It is worth noting that DeepSeek has not released the complete code for independent analysis or benchmarking by third parties, nor has it provided an API for independent testing. The company has not published related blog posts or technical documents explaining the training or construction of R1-Lite-Preview, leaving its origins shrouded in mystery.
R1-Lite-Preview is currently available for free through DeepSeek Chat (chat.deepseek.com), but its advanced "Deep Thought" mode is limited to 50 messages per day, allowing users to experience its powerful capabilities. DeepSeek plans to release open-source versions of the R1 series models and related APIs to further support the development of the open-source AI community.
DeepSeek continues to drive innovation in the open-source AI field, and the release of R1-Lite-Preview adds a new dimension to its capabilities in reasoning and scalability. As businesses and researchers explore applications of reasoning-intensive AI, DeepSeek's open commitment will ensure that its models remain a vital resource for development and innovation.
Official access: https://www.deepseek.com/
Key Points:
🌟 DeepSeek releases the R1-Lite-Preview model, with performance comparable to and exceeding OpenAI's o1.
📊 The model showcases a transparent reasoning process, allowing users to observe logical steps in real time.
🔍 Significant deep learning and logical reasoning capabilities, with plans to release open-source versions and APIs in the future.