Chinese vs Western AI Knowledge Engineering

Two ecosystems, different approaches, shared trajectory
Western Ecosystem
RAG / Agent Platforms

LangChain 100K+

Code-first, composable chains + agents

LlamaIndex 40K+

Data framework for LLM apps

Haystack 18K+

Production NLP pipelines
Agent Frameworks

LangGraph / CrewAI / AutoGen

Code-heavy, developer-oriented
Model Hub

Hugging Face

Global standard for model hosting
vs
Chinese Ecosystem
RAG / Agent Platforms

Dify 136K+

Visual workflow + RAG, MCP support

RAGFlow 75K+

Deep document understanding + OCR

FastGPT 27K+

Enterprise KB, runs on 2GB RAM
Agent Frameworks

Coze Studio (ByteDance) 15K+

Visual, low-code agent building
Model Hub

ModelScope (Alibaba)

Chinese Hugging Face equivalent
Six Key Differences
Code-first, composable libraries
Approach
Visual, low-code, all-in-one
API-first, cloud-native
Deployment
On-premise dominant (data sovereignty)
Discord, Reddit, X/Twitter
Community
WeChat groups, Zhihu, CSDN
More compute available
Innovation Driver
Efficiency under constraints (chip export controls)
Self-regulation emerging
Regulation
748 AI services filed by Dec 2025
Mixed open/closed
Open Source
Open-weight releases outpacing West
Leading Models for Knowledge Engineering

GPT-4o / o1 (OpenAI)

Dominant in API consumption. Closed-weight.

Qwen3 (Alibaba)

235B MoE, Apache 2.0. 256K-1M context.

Claude Opus 4.6 (Anthropic)

1M context. MCP native.

DeepSeek R1

MIT license. Matched o1. Trained for $6M.

Gemini 2.0 (Google)

2M context. Natively multimodal.

Kimi K2.5 (Moonshot)

Agent Swarm: 100 sub-agents, 1,500 tool calls.
Two ecosystems, one trajectory. The Western approach emphasizes composability and developer tools. The Chinese approach emphasizes visual workflows and enterprise deployment. Both are converging on the same destination: the harness is the product.
awesome-llm-knowledge-systems -- Chapter 9: The Chinese AI Ecosystem