Kunlun Wanwei releases Skywork o1 version 4.0 (Skywork o1) and officially opens invitation testing! As the first o1 model in China with Chinese logical reasoning capabilities, Skywork o1 has slow thinking reasoning capabilities. It has built-in thinking, planning, reflection and other abilities in the model output, and its reasoning capabilities on the standard evaluation set have been significantly improved. The three models released this time: Skywork o1Open, Skywork o1Lite and Skywork o1Preview respectively cater to different needs and demonstrate Kunlun’s breakthrough progress in the field of general artificial intelligence, marking another step in its “All in AGI and AIGC” strategy. An important milestone.
Kunlun Wanwei Technology Co., Ltd. announced that the o1 version of its "Skywork Model 4.0" (Skywork o1) has officially launched invitation testing. This news marks an important step for the company in the field of general artificial intelligence, aiming to achieve higher levels of artificial intelligence applications through technological innovation.
Skywork o1 is a series of models released by Kunlun Wanwei Group. It has slow thinking and reasoning capabilities. It is the first o1 model in China with Chinese logical reasoning capabilities. This model not only has the ability to think, plan, reflect, etc. built into the model output, but its reasoning ability on the standard evaluation set is significantly improved compared to ordinary models. The Skywork o1 released this time includes three models: Skywork o1Open, Skywork o1Lite and Skywork o1Preview. Skywork o1Open is an open source model based on Llama3.18B, Skywork o1Lite has complete thinking capabilities, and Skywork o1Preview provides more diversity. and in-depth thought process.
Kunlun Wanwei Group stated that Skywork o1Open has greatly improved both mathematics and code indicators, raising the performance of Llama-3.1-8B to the same niche SOTA level, and unlocking many complex mathematical tasks that cannot be solved by lightweight models. In addition, the company will also open source Process Reward Model (PRM) for two reasoning tasks: Skywork o1Open-PRM-1.5B and Skywork o1Open-PRM-7B. These models can score each step and improve reasoning and thinking capabilities.
The performance improvement of Skywork o1 on logical reasoning tasks is due to Kunlun Wanwei's self-developed training program, which includes reasoning and reflection ability training, reasoning ability reinforcement learning and reasoning planning. The application of these technologies makes Skywork o1 perform well on common sense reasoning, logical reasoning, mathematical reasoning, ethical decision-making and other issues.
Kunlun Technology emphasized that the release of Skywork o1 is not only an important step in the company's "All in AGI and AIGC" strategy, but also an important step in building an AI technology stack. The company will continue to work toward achieving general artificial intelligence, allowing everyone to better shape and express themselves.
Early adopter address: www.tiangong.cn
The editor of Downcodes concluded: Kunlun Wanwei’s Tiangong Model 4.0 o1 version has demonstrated strong capabilities in logical reasoning and complex task processing, providing a new direction for the development of general artificial intelligence. Its open source strategy also provides developers with valuable resources and learning opportunities, and it is worth looking forward to subsequent updates and applications.