Who we are

XingChen-AGI is the open-source organization of China Telecom AI Technology Co., Ltd. We focus on two things:

1. Agent-era large language models that move from passive Q&A to task execution — long-context planning, multi-step reasoning, tool calling, and structured output.
2. Multimodal document intelligence that turns scanned documents, PDFs, tables, formulas and charts into clean, machine-readable data.

Flagship model

Xing4.0-29B-A4B — our agentic flagship

A Mixture-of-Experts (MoE) model — trained end-to-end on a fully in-house stack, from silicon to software.

Benchmarks

• SWE-bench Verified 75.0 — a flagship code-agent score
• SuperCLUE Agent 93.52 (ranked #3)

Native upstream support, merged into the mainlines:

• llama.cpp (#31589, b12174+) — mainline GGUF inference on CPU / CUDA with zero new ggml operators
• SGLang (#39793) — native serving support
• Training: ms-swift (#10275) and LLaMA-Factory (#10754) merged
• vLLM — via the official plugin (Xing4.0-plugin); also integrates with OpenCode, Claude Code
• More: GPTQModel, MLX (mlx-vlm / unsloth-zoo), PocketLLM, LocalAI GGUF gallery, RelicLLM (NPU GGUF)

Models

Xing4.0-29B-Intent — schema-guided structured output

A task-specific model for schema-guided intent recognition and structured information extraction — turns natural language into structured JSON for business systems: intelligent assistants, request routing and document extraction.

• Four tasks in one model: intent recognition & slot filling · named entity recognition · relation extraction · event extraction — unified category + args output, configured by schema
• Top overall score (58) ahead of Gemini-3.6-Flash (49), Qwen3.8-Max (46) and GPT-5.6-Sol (45) — best on 5 of 6 metrics
• Companion dataset: 2.75M samples unifying 61 public datasets in one schema-guided format (4 task types · 19 domains · 52 languages)
• XingIntent SDK — schema-driven training and structured inference tooling

TeleOCR

A compact (~1.2B) document-parsing VLM (formerly NaviDC-OCR) that robustly parses complex layouts, tables, formulas and scientific charts — including distorted real-world captures.

• OmniDocBench V1.6 — #1 with 96.87
• Wild-OmniDocBench — 88.53
• ICDAR 2026 Sci-Chart Document Mining Challenge — Winner

Code and technical report: github.com/caipeng328/TeleOCR · arXiv 2608.12898

Get started

pip install -U huggingface_hub

# download the flagship
huggingface-cli download XingChen-AGI/Xing4.0-29B-A4B

Load it with Transformers:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "XingChen-AGI/Xing4.0-29B-A4B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")

For serving, quantization and hardware requirements, see the model card of each build. Or try it on the Tianyi AI open platform.

Where to find us

Built by China Telecom Artificial Intelligence Technology Co., Ltd.