China Telecom has open-sourced the TeleChat-7B version of the Xingchen Semantic Large Model, providing a 1T cleaned dataset. They also plan to release the 12B version of the model on January 20th, aiming to collaborate with more developers in building an open-source large model ecosystem. The Xingchen Semantic Large Model is a language model trained on 1.5 trillion tokens of Chinese and English corpora. It introduces four major techniques that reduce the hallucination rate of AI large models by 40%, enabling better contextual understanding and providing relevant answers. The Xingchen Semantic Large Model has been applied in various scenarios within and outside China Telecom, including writing, code programming, network fault analysis, enterprise management analysis, government consultation, and public service reception. It has achieved high levels of generated word count, adoption rate, coverage, and accuracy of understanding. China Telecom unveiled the Xingchen Semantic Large Model at the 2023 Digital Technology Ecology Conference and announced the schedule for its open-source release. The TeleChat-7B version includes the TeleChat-7B-bot dialogue model, huggingface format weight files, as well as int8 and int4 quantized versions of the 7B model. The Xingchen Semantic Large Model has been adapted for Huawei Ascend AI hardware and software, supporting the Atlas 300I Pro inference card and Atlas training server. Users can train and infer models using the MindSpore and PyTorch frameworks. Github open-source address: https://github.com/Tele-AI/Telechat Gitee address: https://gitee.com/Tele-AI/tele-chat