Chinese AI tools: a swarm of agents, a big context, and what it costs

For owners and executives looking for affordable alternatives to Western services. After this article you will understand how the Chinese AI ecosystem is put together, how a “swarm of agents” works on research tasks, and what to decide before you upload commercial data into it.

Your team needs to enter a new market. That means studying the competitors, collecting prices, reading through two dozen PDF reports, and assembling a sales strategy out of all of it. A competent analyst will spend several days on it.

An ordinary LLM (a large language model trained on large bodies of text) only helps part of the way here. It runs into either speed or the volume of information it can hold “in its head” at once. So instead of analysis you get a set of fragments that you still assemble by hand.

In AKORDO’s practice we see the choice usually sit between an expensive subscription to a Western service, which can also be blocked, and consolidating the data by hand. Chinese developers are building a third option: not a single model but an ecosystem, where AI works less like an information desk and more like a group of specialist employees inside your working environment.

A swarm of agents: when one task is split between five

The concept is called Agent Swarm. An agent, unlike a chatbot, is an autonomous program that carries out a sequence of actions on its own to reach a goal it has been set. Swarm means several such programs start up, and each does its own part.

In Kimi this exists as a mode where several assistants work on one request in parallel. The first searches for sources on the internet. The second checks the competitors. The third pulls what has been found into a table. The fourth looks at the task through the eyes of a potential client or a skeptic. The fifth gathers all of it into a final report.

The approach rests on the “wisdom of the crowd”: several independent views of one idea give a soberer assessment than a single answer. What comes out is not a paragraph of text but a folder of analysis, put together in a few minutes.

It gives you most where the documents are many and mixed: PDF reports, audio recordings of meetings, competitors’ pages. One handler chokes on that volume, a distributed group goes through it calmly. We recommend putting preparation for a product launch and an audit of current processes on this, the tasks where shallow analysis is expensive.

A million tokens of memory

The second familiar problem in working with AI is called context. This is the model’s working memory, the space in which it holds the details of your conversation. As the dialogue gets longer, an ordinary model starts forgetting the original conditions, and errors appear in the report that were not there at the start.

The Qwen model works with a context window of up to a million tokens (units of information). In practice that changes three things.

You upload a whole project folder and hold the conversation inside it without restating the input data every time. You analyze an annual report or six months of correspondence with a client, and the model does not lose details from the beginning. And you can ask for “artifacts” in the chat, meaning instant prototypes of pages or landing pages.

For a commercial director that means one working environment holding the sales rules, the scripts, and the history of the relationship with key partners, with AI behaving like an assistant who has read the material.

The bet on low cost

The Western market mostly competes over the most powerful model. The Chinese one builds a different strategy: volume and a low cost per operation. The reason is prosaic. Limited access to high-performance chips forces them to win on efficiency rather than on the power of the hardware.

DeepSeek became the symbol of this approach by showing that quality text processing in expert modes costs noticeably less than the market equivalents.

What that gives you in practice. Automation reaches processes where the price of a request used to make it uneconomic. Free tools appear for “vibe coding,” the fast assembly of application prototypes and presentations without knowing how to program. And access to the models themselves runs through aggregators, which takes registration and payment off the table for users outside China.

For a small business with no budget for IT infrastructure, that is the main argument.

Video through a storyboard

Generating video and images usually gives an unpredictable result: you spend tokens on ten attempts and one of them fits. The way to take the process under control is to split it into two steps.

First you create the storyboard, a static image with the sequence of scenes for the clip to come. Then a specialist model (Seedance 2.0 in the source’s example) generates the video from that storyboard.

You see the structure of the clip before you have paid for it, and the model does not improvise. For marketing this is a way to make advertising mockups and presentation clips without a production studio, spending the budget once rather than on every attempt to guess right.

Is this about you?

Check yourself against this list:

• blocks or the cost of Western AI services already limit your work;
• your people regularly have to analyze mixed documents (PDF, audio, websites) in one flow;
• you want to automate routine requests and have nothing to build your own system with;
• marketing needs frequent visual content: video, images, presentations;
• you lack a “second opinion” on strategy, a view from the client’s side or from a skeptic’s.

Three matches or more mean a week is worth setting aside for a test.

Where to start

Take one task and one model. Upload your current sales scripts into a model with a large context and ask it to find the contradictions in them, for example. It is a quick check with a measurable result.

Before that, settle a separate question none of these services will close for you: which data exactly you are prepared to hand to somebody else’s cloud. Sales scripts, a client base, and financial statements carry different costs if they leak, and the decision here belongs to the owner, not to the tool. Start with data whose loss would not trouble you.

Do not move every process at once. Pick a service with a clear interface that needs no configuration, and judge the quality of the result while the investment is still measured in hours.

Key takeaways

• Chinese AI is built as an integrated ecosystem betting on affordability and volume rather than on the most powerful model.
• A swarm of agents breaks a complex task into parallel roles and imitates the work of a whole department in a few minutes.
• A million-token context removes the model’s “forgetfulness” problem on long projects.
• A storyboard before video generation gives control over the result back and saves the marketing budget.

The AKORDO team built sales and operations analytics in three cases: AI Data Intelligence for anomaly detection at a fintech company, regular KPI analysis and management reporting in B2B manufacturing and a leadership dashboard for e-commerce sales. For AI analysis of calls and chats, read the guide to call quality control. To choose the first task for an AI agent, read the guide to the first AI agent scenario.