Two groundbreaking Chinese AI models, DeepSeek-R1 and Kimi k1.5, have caused a sensation in Silicon Valley, with their performance rivaling OpenAI’s o1 model. These models have captured the attention of the global tech community, signaling a significant milestone in China’s AI journey.
Model Performance and Features
The Kimi k1.5 model has achieved, and in some aspects surpassed, the capabilities of OpenAI’s o1 in terms of mathematical, coding, and multi-modal reasoning under the long-CoT framework. This is the first time a company has reached such a level globally.
Technological Approach Differences
–
DeepSeek-R1 adopts the AlphaZero method, self-directing through reinforcement learning and human input without any supervised fine-tuning.
–
Kimi k1.5 utilizes the AlphaGo-Master approach, undergoing lightweight SFT preheating through prompted engineering of CoT (Chain of Thought) trajectories.
The Next Step for Domestic o1
The difference between the two models lies in user experience. While DeepSeek-R1 is a text reasoning model, Kimi k1.5 supports text and visual multi-modal reasoning. Multi-modality is a necessary path to achieving AGI, and Kimi has already unveiled its roadmap for the k series of thinking models.
Shaking the OpenAI Myth
This breakthrough in Chinese AI has begun to shake the dominant position of OpenAI. The progress of Chinese AI has been reported by overseas communities and American tech media, with OpenAI’s technical halo and dominance seemingly diminishing.
US Response
The US government has also made moves in the AI field, announcing a project led by OpenAI, in collaboration with SoftBank and other giants, to build a massive AI infrastructure center. This project is being viewed as the AI era’s Manhattan Project.
Below is the formatted content for a WordPress blog post:
Two Chinese AI models, DeepSeek-R1 and Kimi k1.5, have caused a stir in Silicon Valley, with their capabilities matching those of OpenAI’s o1 model. DeepSeek employs reinforcement learning, while Kimi utilizes lightweight SFT preheating. The multi-modal reasoning abilities of these models are particularly impressive, with Kimi k1.5 surpassing OpenAI’s model in certain areas. These breakthroughs challenge OpenAI’s status and have sparked heated discussions in overseas communities. In response, the US has announced new AI infrastructure plans, reflecting recognition and a challenge to China’s AI advancements.
