Today, Kunlun Tech Group officially announced that its independently developed Tiangong Large Model 4.0 series - o1 and 4o versions have been launched simultaneously, covering both web and app platforms to provide advanced AI services for free to a wide range of users.
Kunlun Tech Group introduced that the Tiangong Large Model 4.0 o1 version, as the first domestically produced model with Chinese logical reasoning capabilities, not only offers the convenience of immediate open-source upon release but also launches two specialized versions with superior performance. This model demonstrates exceptional abilities in handling various reasoning challenges such as mathematics, coding, logic, common sense, and ethical decision-making, providing more intelligent and efficient solutions for users.
At the same time, the Tiangong Large Model 4.0 4o version, a multimodal model, has also attracted significant attention. Kunlun Tech Group has launched Skyo, a real-time voice dialogue assistant powered by this model. Skyo offers unparalleled intelligent voice dialogue experiences with its unique emotional expression, quick response times, and seamless multilingual switching capabilities.
When discussing the core technologies of the Tiangong Large Model 4.0 series, Kunlun Tech Group revealed the three-stage self-research training plan for Skywork o1. Initially, they constructed high-quality step-by-step thinking data through a self-developed multi-agent system to continue pre-training and supervised fine-tuning of the base model, significantly enhancing its reasoning and reflective abilities. Secondly, the team developed the latest Skywork o1 Process Reward Model (PRM) tailored for step-by-step reasoning enhancement, combined with a proprietary step-by-step reasoning strengthening algorithm, further boosting the model's reasoning and thinking capacities. Finally, based on Kunlun's self-developed Q online reasoning algorithm, the model can think online and find the optimal reasoning path. The realization and public release of this innovative technology mark the world's first application of the Q algorithm in online reasoning, greatly improving the model's efficiency.
