TeleChat is an open-source large language model family on GitHub developed by China Telecom’s AI company (Tele-AI). The repository, with about 1,900 stars, releases the TeleChat chat models in 1B, 7B and 12B sizes together with quantized versions, training details and a technical report.
TeleChat models are trained on trillions of high-quality Chinese and English tokens: the 7B base model used roughly 1.5 trillion tokens, while the 12B model scaled to about 3 trillion with improvements to architecture, data and training methods that substantially raised its scores on general question answering, knowledge, code and math benchmarks. Alongside the main chat models, the team open-sourced int8 and int4 quantized variants for cheaper deployment, plus an open trillion-token Chinese training dataset. Weights are distributed through Hugging Face and ModelScope in standard formats, and inference is supported through frameworks such as vLLM, making TeleChat a practical open foundation for Chinese-language applications.
TeleChat is part of China Telecom’s broader Xingchen (TeleAI) artificial intelligence program, and the open releases are intended to support domestic and international developers building on Chinese open models, including adaptations for domestic Chinese hardware platforms documented in the ecosystem.
TeleChat is open source and free to download and self-host from GitHub, Hugging Face or ModelScope. There are no API fees or subscriptions; users provide their own compute for training and inference.
TeleChat is for developers, researchers and enterprises building Chinese-language LLM applications who want a capable open model family with multiple sizes and quantization options that they can fine-tune and deploy on their own infrastructure.