Xing4.0-29B-A4B

From "answering questions" to "getting things done" — an agentic MoE flagship
with 29B total / 4B active parameters and a native 256K context.

🦉 GitHub • 🤗 Hugging Face Org • 🤖 ModelScope • 🪐 Modelers • ⚡ vLLM Plugin • ☁️ Tianyi AI Platform

Pick a build:  BF16 (this repo) • FP8 • GGUF (2–8-bit, 11–34 GB, runs in llama.cpp) • W8A8

---

Contents

• Introduction
• Model at a Glance
• Highlights
• Architecture
• Benchmarks
• Quickstart
  • Inference
  • Serving & Ecosystem
• Recommended Parameters
• Evaluation Details
• Citation
• Acknowledgments

Introduction

Xing4.0-29B-A4B is a next-generation large language model in the Xing series (formerly TeleChat), developed by China Telecom Artificial Intelligence Technology Co., Ltd. With 29B total parameters and only 4B activated per token, it natively supports a 256K context length, extensible to 512K. It was trained end-to-end on a fully in-house stack — from silicon to software — and is deeply optimized for complex engineering tasks: long-context planning, multi-step reasoning, tool calling, and structured output.

For full project documentation, deployment guides, and fine-tuning recipes, visit the GitHub repository.

Model at a Glance

Highlights

• Agent-oriented architecture — built on the mHC + MLA + MTP architecture, supporting multi-step planning, tool calling, and complex reasoning-chain execution, ensuring task coherence and execution stability under long contexts.
• Deep software–hardware co-optimization — adapted to domestic NPU clusters with a fully in-house training framework, including mHC feature adaptation and fused-operator development, enabling stable and efficient training at scale.
• ~96% training throughput gain — fine-grained MoE communication optimization, selective recomputation, automatic graph–operator fusion, and hand-tuned fused mHC kernels lifted overall training throughput by approximately 96% over out-of-the-box performance.
• Native in the open ecosystem — fine-tuning via ms-swift and LLaMA-Factory merged upstream; serving via SGLang mainline (#39793) and the official vLLM plugin; local GGUF inference via llama.cpp mainline (#31589, b12174+); format-aligned with agent frameworks such as OpenCode, Claude Code, OpenClaw, and Hermes for out-of-the-box workflows.
• Easy adaptation for domain scenarios — lightweight customization on proprietary data for intent classification, table understanding, contract auditing, and knowledge-based QA, for rapid and low-cost domain capability deployment.

Architecture

Benchmarks

Head-to-head with same-class open agentic models:

Bold marks the best score in each benchmark.

• Coding agent — 75.0 on SWE-bench Verified (approaching Qwen3.6-35B-A3B's 76.0, far ahead of Gemma4-26B-A4B's 53.0), 66.0 on SWE-bench Multilingual, and a clear lead on Terminal-Bench 2.1 (57.5 vs 51.5 / 30.0).
• General agent — best in this comparison on Claw-Eval (76.55) and DeepresearchBII (60.80), with 64.63 on Tau3-Bench.

Quickstart

Inference

Xing4.0-29B-A4B can be accessed via an OpenAI-compatible API.

from openai import OpenAI

client = OpenAI(
    base_url="your-base-url",
    api_key="your-api-key",
)

completion = client.chat.completions.create(
    model="Xing4.0-29B-A4B",
    messages=[{"role": "user", "content": "Briefly explain the basic principles of quantum computing."}],
    temperature=1.0,
    top_p=0.95,
    extra_body={
        "repetition_penalty": 1.05,
        "skip_special_tokens": False,
        "spaces_between_special_tokens": False,
        "chat_template_kwargs": {
            "enable_thinking": True,  # Set to False to disable thinking
        },
    },
)

print(completion.choices[0].message.content)

Serving & Ecosystem

Xing4.0-29B-A4B runs across serving stacks with different integration levels (as of 2026-11-10):

Prebuilt container images and step-by-step deployment guides are listed in the GitHub repository.

Recommended Parameters

Evaluation Details

• SWE-bench Verified: We evaluate Xing4.0-29B-A4B using the SWE-agent harness with temperature=1.0, top_p=0.95, repetition_penalty=1.05 and a 210K context window.
• Terminal-Bench 2.1: We evaluate in terminus-2 with temperature=0.8, top_p=0.95, repetition_penalty=1.05, max_tokens=64K and a 24-hour timeout. We report the average score over 3 runs.
• SWE-bench Multilingual: We evaluate Xing4.0-29B-A4B using the SWE-agent harness with temperature=1.0, top_p=0.95, repetition_penalty=1.05 and a 210K context window.
• Claw-Eval: We evaluate using the official harness with temperature=0.8, top_p=0.95, repetition_penalty=1.05, max_tokens=16384 and a 256K context window. We report the average score over 3 runs.
• Tau3-Bench: We evaluate using the official harness with temperature=0.8, top_p=0.95, repetition_penalty=1.05, max_tokens=16384. We report the average pass^1 over 4 runs.
• DeepresearchBII: We evaluate using the OpenCode harness with temperature=0.8, top_p=0.95, repetition_penalty=1.05, max_tokens=32768 and a 256K context window, with Exa MCP server enabled.
• AIME2026: We evaluate with temperature=1.0, top_p=0.95, repetition_penalty=1.05, max_tokens=131072. We report the average score over 5 independent runs.
• IFBench: We evaluate with temperature=0.8, top_p=0.95, repetition_penalty=1.05, max_tokens=81920.
• AA.LCR: We evaluate with temperature=1.0, top_p=0.95, repetition_penalty=1.05, max_tokens=65536. We report the average score over 3 runs.

Citation

@misc{liu2025trainingreporttelechat3moe,
      title={Training Report of TeleChat3-MoE}, 
      author={Xinzhang Liu and Chao Wang and Zhihao Yang and Zhuo Jiang and Xuncheng Zhao and Haoran Wang and Lei Li and Dongdong He and Luobin Liu and Kaizhe Yuan and Han Gao and Zihan Wang and Yitong Yao and Sishi Xiong and Wenmin Deng and Haowei He and Kaidong Yu and Yu Zhao and Ruiyu Fang and Yuhao Jiang and Yingyan Li and Xiaohui Hu and Xi Yu and Jingqi Li and Yanwei Liu and Qingli Li and Xinyu Shi and Junhao Niu and Chengnuo Huang and Yao Xiao and Ruiwen Wang and Fengkai Li and Luwen Pu and Kaipeng Jia and Fubei Yao and Yuyao Huang and Xuewei He and Zhuoru Jiang and Ruiting Song and Rui Xue and Qiyi Xie and Jie Zhang and Zilu Huang and Zhaoxi Zhang and Zhilong Lu and Yanhan Zhang and Yin Zhang and Yanlei Xue and Zhu Yuan and Teng Su and Xin Jiang and Shuangyong Song and Yongxiang Li and Xuelong Li},
      year={2025},
      eprint={2512.24157},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2512.24157}, 
}

@misc{wang2025technicalreporttelechat2telechat25,
      title={Technical Report of TeleChat2, TeleChat2.5 and T1}, 
      author={Zihan Wang and Xinzhang Liu and Yitong Yao and Chao Wang and Yu Zhao and Zhihao Yang and Wenmin Deng and Kaipeng Jia and Jiaxin Peng and Yuyao Huang and Sishi Xiong and Zhuo Jiang and Kaidong Yu and Xiaohui Hu and Fubei Yao and Ruiyu Fang and Zhuoru Jiang and Ruiting Song and Qiyi Xie and Rui Xue and Xuewei He and Yanlei Xue and Zhu Yuan and Zhaoxi Zhang and Zilu Huang and Shiquan Wang and Xin Wang and Hanming Wu and Mingyuan Wang and Xufeng Zhan and Yuhan Sun and Zhaohu Xing and Yuhao Jiang and Bingkai Yang and Shuangyong Song and Yongxiang Li and Zhongjiang He and Xuelong Li},
      year={2025},
      eprint={2507.18013},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2507.18013}, 
}

Acknowledgments

Xing4.0-29B-A4B is better because of the open-source community. This section is our way of saying thanks — and we will keep adding names as the community grows.

Inference, training, and tooling. Thank you to the maintainers of vLLM, SGLang, KTransformers, TensorRT-LLM, and llama.cpp, and to the wider MLX ecosystem — for building inference stacks open enough to carry an architecture combination like ours, and for keeping them that way. Support for Xing4.0 has also landed in ms-swift, GPTQModel, and unsloth-zoo, and in LocalAI's gallery. Thank you to their maintainers.

Community quants and ports. Special thanks to the community members who did real work on this model and published it for everyone:

• @kingjones777 — AMD ROCm builds with public recipes and hashes, and detailed reports that made the ROCm path work end to end.
• Venastine Research — four GGUF tiers, including an 11.6 GB build that puts the model within reach of 16 GB cards, with a fully documented imatrix recipe.
• @jmarceno — an independently recomputed imatrix and a 3.31 BPW quant, with the recipe and hashes published.
• @TokenAI-zer — an MLX implementation verified to near-exact logit parity, and precise issue reports on the reference implementation.
• @nativ-community — an MLX 8-bit port with token-level fidelity verification, with upstream work now merged in mlx-vlm.
• The FlagOS project — multi-accelerator builds of the model.

We are also grateful to the teams and individuals who fine-tuned Xing4.0 for domain work — finance, official document writing, roleplay, meeting minutes, rewriting, information extraction — and published those weights for others to build on.

If you have ported, quantized, evaluated, or fixed Xing4.0 and are not on this list — thank you. Open a discussion and let us know; we will add you.

Evaluation tooling. Our benchmark numbers rest on harnesses built and maintained by the SWE-agent, Terminal-Bench, tau-bench, and Claw-Eval teams, along with research tooling from Exa. We are grateful for the infrastructure that makes open model evaluation possible.

---

China Telecom Artificial Intelligence Technology Co., Ltd.
More open models from XingChen-AGI