AI BriefingChina Telecom releases Xing4.0-29B-A4B, a 29B MoE model trained on Ascend NPUs
·2026.11.07 04:19
Key point
The model achieves 75.00 on SWE-bench Verified and 90.00 on AIME2026, supporting 256K context natively.
Details
- China Telecom Artificial Intelligence Technology Co., Ltd. released Xing4.0-29B-A4B, a 29B parameter Mixture-of-Experts (MoE) model with 4B active parameters per token.
- The model is the first of its scale trained entirely on the Ascend NPU platform using the MindSpore framework, achieving approximately 96% improvement in training throughput over out-of-the-box performance.
- It natively supports a 256K context length, extensible to 512K, and utilizes an mHC + MLA + MTP architecture for agent-oriented tasks like multi-step planning and tool calling.
- Benchmark results show 75.00 on SWE-bench Verified, 90.00 on AIME2026, and 69.67 on IFBench, outperforming Gemma4-26B-A4B and Qwen3.6-35B-A3B on several coding and reasoning metrics.
- Full open-source ecosystem compatibility is provided, supporting fine-tuning via LLaMA-Factory and MindFormers, and inference via SGLang, vLLM, and KTransformers.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Craft Anything: Open-source Rust engine for AI-driven software reverse engineering and recreation