China AI Hub
A2A Protocol: How Chinese AI Agents Talk to Each Other
Agent2Agent (A2A) is an open protocol announced by Google in April 2025 that lets AI agents from different vendors and frameworks discover, communicate and coll…
Technology Knowledge Base
Core AI technologies — how they work, why they matter, how China's companies implement them, and their limitations.
AI Agents in China: Frameworks, Vendors and Limits
An AI agent is a system in which a language model acts autonomously toward a goal — planning, calling tools, observing results and iterating — rather than produ…
AI Chips in China: Ascend, Domestic GPUs and Supply
AI chips are specialized processors — GPUs and purpose-built accelerators — designed for the matrix-heavy compute of model training and inference, the physical …
AI Infrastructure in China: Clusters, Power, Constraints
AI infrastructure is the systems layer beneath models — compute clusters, storage, networking, and the serving software — that turns raw hardware into a service…
Computer-Use Agents in China: Capability and Risk
Computer use is a model capability in which the model operates a graphical computer the way a person does — viewing the screen, moving the mouse and typing — en…
Function Calling in Chinese Models: Support and Gaps
Function calling is OpenAI's term for tool calling: a model capability in which the model produces a structured request (name plus arguments) to invoke a develo…
Deep Research Agents in China: Products and Gaps
Deep research is an agentic workflow in which a model autonomously plans a research question, performs many rounds of search and document reading, and produces …
Long Context in Chinese Models: 1M Tokens in Practice
Long context refers to a model's ability to accept very large inputs — hundreds of thousands to millions of tokens — in a single request, allowing whole codebas…
Inference in China: Serving Costs and Optimisation
Inference is the process of running a trained model to generate outputs — and, as a discipline, the engineering of making that process fast and cheap through at…
Distillation in Chinese AI: Method, Risk and Detection
Knowledge distillation is a training technique in which a smaller 'student' model learns to imitate the outputs of a larger 'teacher' model, transferring capabi…
MCP in Chinese AI: Adoption, Servers and Limitations
Model Context Protocol (MCP) is an open protocol, introduced by Anthropic in 2024, that standardizes how applications expose tools, resources and context to AI …
Mixture of Experts in Chinese Models: Activation Explained
Mixture of Experts (MoE) is a model architecture where each layer contains many small 'expert' networks and a router activates only a few per token, so total pa…
Multimodal AI in China: Vision, Video and Audio
Multimodal AI refers to models that process more than one input type — text, images, audio, video — in a unified model, enabling tasks like image understanding,…
Quantization for Chinese Open Models: FP8, INT4, Quality
Quantization compresses model weights and activations from high-precision formats (FP16/BF16) to lower precision (8-bit, 4-bit or mixed), shrinking memory and c…
RAG in Chinese AI Stacks: Retrieval, Vendors and Pitfalls
Retrieval-Augmented Generation (RAG) is a pattern where a language model's answers are grounded in documents fetched by a retrieval system at query time, combin…
Reasoning Models in China: Test-Time Compute and Cost
A reasoning model is a language model trained to produce explicit intermediate reasoning — typically chain-of-thought steps — before giving a final answer, usua…
Synthetic Data in Chinese AI: Use, Risk and Disclosure
Synthetic data is training data generated by models rather than collected from human sources — used to expand scarce domains, distill capabilities, and build re…
Tool Calling in Chinese Models: What Actually Works
Tool calling is a model capability where the model, instead of only writing text, emits structured requests to invoke external tools — search, calculators, code…