Xing4.0
An open-weight, agent-oriented language model with 29 billion total parameters and 4 billion active per token, trained entirely on Ascend NPU hardware with a native 256K-token context window.
Xing4.0-29B-A4B is developed by China Telecom Artificial Intelligence Technology and is the latest generation of the TeleChat model line, built on an architecture the developers call mHC combined with multi-head latent attention and multi-token prediction, aimed at multi-step planning, tool calling and long, coherent reasoning chains. It is, according to its authors, the first model at this scale trained entirely on Ascend 910C NPU clusters using the MindSpore and MindFormers frameworks, with custom fused operators and cross-framework numerical alignment work that the README credits with a large training-throughput improvement over an out-of-the-box setup.
The model supports fine-tuning through LLaMA-Factory and MindFormers, inference and deployment through SGLang, vLLM and KTransformers, and integration with agent frameworks including OpenCode, Claude Code, OpenClaw and Hermes, with multi-chip deployment support through BAAI FlagOS.
It is licensed under Apache-2.0.