DeepSeek Launches R1-Lite-Preview, Outperforms OpenAI’s o1 Model

DeepSeek, a Chinese AI research lab backed by High-Flyer Capital Management, has released the DeepSeek-R1-Lite-Preview, a reasoning AI model built to challenge OpenAI’s o1 model.

The model’s performance is reportedly on par with OpenAI’s o1-preview on rigorous benchmarks such as AIME and MATH, which test LLMs logical and mathematical reasoning skills. 

The DeepSeek-R1-Lite-Preview model introduces “chain-of-thought” reasoning, providing users with a detailed step-by-step explanation of its problem-solving process. This feature addresses a common criticism of AI models—lack of transparency—by allowing users to understand the reasoning behind the model's conclusions. 

DeepSeek-R1-Lite-Preview reveals Inference Scaling Laws: Longer Reasoning Results in Better Performance. The company reported that the model shows steady improvements in AIME scores with increased reasoning length.

The introduction of DeepSeek-R1-Lite-Preview comes amid growing scrutiny of traditional AI scaling laws, which suggest that increasing data and computational power will continuously improve model capabilities. Instead, DeepSeek employs test-time compute techniques, allowing the model additional processing time during inference to tackle complex tasks more effectively.

DeepSeek’s new model is available through its web-based chatbot, DeepSeek Chat, where users can experience the model’s capabilities firsthand. However, usage is currently limited to 50 messages per day. 

Despite its impressive performance, the model faces challenges typical of AI systems developed in China, including restrictions on politically sensitive topics due to regulatory pressures.

DeepSeek plans to release open-source versions of its R1 models and associated APIs soon, reinforcing its commitment to transparency and accessibility in AI development. This move is expected to intensify competition among major Chinese tech companies like ByteDance, Alibaba and Baidu.

Alibaba recently launched the Qwen2.5-Turbo, with an expanded context length of 1M, roughly 1 million English words or 1.5 million Chinese characters—equivalent to 10 novels, 150 hours of speech, or 30,000 lines of code.