TeleChat3
A family of large language models trained entirely on domestic Chinese computing infrastructure by China Telecom's AI research institute, released with reasoning-focused variants.
TeleChat3 is a large language model series developed by the AI Research Institute of China Telecom, trained entirely on domestically produced computing hardware. The initial release includes a 105-billion-parameter Mixture-of-Experts model with about 4.7 billion active parameters and a dense 36-billion-parameter model, both in a reasoning-focused thinking configuration, with model weights published on Hugging Face-compatible repositories and ModelScope.
The repository documents the model architecture, including layer counts, hidden sizes and attention type for each variant, and reports benchmark scores across knowledge, mathematics, creative writing, instruction following, code generation and agentic tool use compared with several other open models. It supports local inference across single and multiple GPUs, with documented deployment paths through vLLM and SGLang, and fine-tuning through LLaMA-Factory.