🚀 Our Models

Xing4.0-29B-A4B

Our flagship 29B MoE language model with only 4B activated parameters:

• 🧠 256K context natively, extensible to 512K
• 🏆 75.0 on SWE-bench Verified · 93.52 on SuperCLUE-Agent
• 💻 4-bit quantization runs locally on a single GPU with ~15GB VRAM (RTX 3090/4090)
• ⚙️ Official runtimes: vLLM, SGLang, TensorRT-LLM, llama.cpp, KTransformers
• 📦 Precision builds: FP8 · GGUF · W8A8

Xing4.0-29B-Intent

A task-specific model built on Xing4.0-29B-A4B 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) on a 6,000-sample benchmark drawn from 48 public datasets — best on 5 of 6 metrics
• 🌐 Chinese & English
• 📦 Precision build: FP8
• 📚 Companion dataset — 2.75M samples unifying 61 public datasets in one schema-guided format across 4 task types, 19 domains and 52 languages, plus China Telecom's in-house smart-device assistant annotations (Apache-2.0, upstream licenses preserved)
• 🧰 XingIntent SDK — open-source tooling for schema-driven training and structured inference

TeleOCR

A compact multimodal model that turns scans, PDFs and photos into structured text — complex layouts, tables, formulas and charts included.

• 🏆 96.87 on OmniDocBench V1.6 · 88.53 on Wild-OmniDocBench
• 🌐 Bilingual (Chinese & English)

🛠 Trained in-house, runs everywhere

All models are trained in-house on 1.5T tokens of curated bilingual data, and run across a wide range of hardware — from a single consumer GPU to production clusters.

🤝 Community

• 30+ community derivatives — GGUF/MLX/ROCm quantizations and multi-chip adaptations (incl. BAAI FlagOS)
• Questions, deployment help or derivative showcases — open a Discussion on any repo, we respond.
• Code & mirrors: GitHub · ModelScope · Gitee

Named to Hugging Face's trending board top 4 (Sep 2026).

Maintained by the Xingchen General Artificial Intelligence Laboratory, China Telecom Artificial Intelligence Technology Co., Ltd.

_{Benchmarks are internal test results and do not represent final product quality.}