New Large Language Model Released by China Telecom AI

China Telecom Artificial Intelligence Technology Co., Ltd. has released Xing4.0-29B-A4B, a new large language model (LLM) within its Xing series, formerly known as TeleChat. According to documentation provided via GitHub and Hugging Face, the model utilizes a Mixture-of-Experts (MoE) architecture, featuring 29 billion total parameters with 4 billion activated per token.

The model is designed for complex engineering and agent-oriented tasks. It utilizes an architecture comprising mHC, Multi-Head Latent Attention (MLA), and Multi-Token Prediction (MTP) to support multi-step planning and tool calling. Notably, the model natively supports a context length of 256K tokens, which is extensible to 512K.

Hardware Optimization and Training Efficiency

The Xing4.0-29B-A4B is the first model of its scale trained entirely on the Ascend NPU platform using the MindSpore framework. Technical specifications indicate the model was adapted for Ascend 910C clusters.

The developers report significant gains in training efficiency through various co-optimization techniques, including fine-grained MoE communication optimization and selective recomputation. According to the developer’s technical highlights, these methods resulted in an approximate 96% increase in training throughput compared to out-of-the-box performance.

The model is compatible with several existing ecosystems, including LLaMA-Factory and MindFormers for fine-tuning, as well as vLLM and SGLang for inference. It has also been adapted for integration into agent frameworks such as OpenCode and Claude Code.

Benchmark Performance

In comparative testing, the Xing4.0-29B-A4B demonstrated competitive performance against other models in specific categories. According to benchmark data released by the developers, the model achieved the following results:

  • Claw-Eval: 76.55
  • SWE-bench Verified: 75.00
  • Terminal-Bench 2.1: 57.50
  • DeepresearchBII: 60.80

When measured against other models, the Xing4.0-29B-A4B outperformed Gemma4-26B-A4B in the Claw-Eval, SWE-bench Verified, Terminal-Bench 2.1, and DeepresearchBII metrics. However, it trailed behind Qwen3.6-35B-A3B in the AIME2026 and Tau3-Bench benchmarks.

Industry Context

The release of the Xing series comes amid broader speculation regarding upcoming model releases from other major AI developers. While unverified reports on social media platforms like Reddit have suggested that Anthropic may be developing a “Model 4” capable of scoring above 50% on various capability benchmarks, and rumors circulate regarding a potential release of ChatGPT Astra 6.2, these claims lack independent confirmation.