The official WeChat channel of the China Telecom AI Research Institute proudly declares that its TeleAI team has successfully overcome technical challenges and completed China's first ultra-large-scale AI model project with trillion parameters, trained on a fully domestically produced WanCard cluster. Additionally, they have historically released the first 100-billion-parameter semantic AI model, TeleChat2-115B, built entirely with indigenous WanCard clusters and a self-developed deep learning framework. This milestone signifies China's substantial progress in developing AI large-scale models with domestic independence, ensuring safety and controllability.

This milestone achievement not only showcases the research prowess of the China Telecom AI Research Institute but also represents another significant outcome of China's strategy towards technological self-reliance and strength. The creation of TeleChat2-115B signifies a fundamental shift in domestic large-scale model training technology, moving from dependence on imports to full domestic production, thereby establishing a robust foundation for the security and growth of China's AI industry.

According to official sources, TeleChat2-115B was trained using the powerful support of China Telecom's self-developed Tianyi Cloud "Integrated Intelligence Computing Service Platform" and "Xinghai AI Platform." The team implemented a series of innovative optimization strategies, achieving GPU computational efficiency exceeding 93% while maintaining model training accuracy. Furthermore, the effective training duration accounted for over 98%, significantly enhancing both the efficiency and stability of the training process.

To address the challenges of training ultra-large parameter models, TeleAI innovatively employed the "small model ensemble" approach. By deploying and validating various model architectures on a large scale, and integrating precise data allocation strategies, they utilized regression prediction models to optimize data configuration. This resulted in efficient resource utilization and a significant enhancement in model performance.

During the model post-training phase, the TeleAI team pursued excellence by generating a vast amount of Q&A data tailored to specific domains such as mathematics, coding, and logical reasoning. They also adopted advanced Supervised Fine-Tuning (SFT) techniques and an iterative update strategy to continuously improve the model's ability to handle complex instructions and enhance the diversity and accuracy of its responses. Through a combination of model synthesis, manual annotation, and rejection sampling, they successfully acquired high-quality SFT training data and representative Reward Model (RM) data, laying a solid foundation for the ongoing optimization of the model's performance.