NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

A Chinese AI Firm Makes 91% of Revenue Overseas — and Pivots Its Products to Agents

Kunlun Wanwei's half-year results show a shift from model benchmarks to task-executing software, with most sales coming from outside China.

2026-11-22 · 913 words · NeuroAI
A Chinese AI Firm Makes 91% of Revenue Overseas — and Pivots Its Products to Agents

The product on the screen is not a chatbot. It reads a folder of documents, builds a spreadsheet, drafts a slide deck, and in one version generates a video to go with them. The user describes the outcome and checks the work. This is the category Kunlun Wanwei (昆仑万维) has spent 2026 pushing: the AI agent (智能体) as a production tool rather than a conversational companion.

The company is an unusual case in China's AI industry. It started as a web game publisher, listed on Shenzhen's ChiNext board in 2015, bought into overseas platforms including Opera, and then rebuilt itself around generative AI. Its half-year 2026 revenue was ¥5.359 billion (≈ US$755M), up 43.55% year on year, according to China Securities Journal. The company says overseas markets account for 91% of revenue.

Key takeaways

  • Kunlun Wanwei reported first-half 2026 revenue of ¥5.359 billion (≈ US$755M), an increase of 43.55% year on year.
  • The company states that overseas revenue accounts for 91% of the total.
  • Its flagship product line is Skywork Super Agents (天工超级智能体), updated several times during 2026.
  • Other products include SkyReels and DramaWave for short dramas (短剧), Mureka for music, and Linky for social interaction.
  • Benchmark claims about its models come from the company; the financial figures come from a securities filing.

From chat to execution

The shift is visible in the release sequence during 2026. In January, the company added Skywork Video v1.0, putting video generation inside the agent workflow. In February it shipped a Windows desktop version that reads and processes local files without uploading them to the cloud, and a cloud agent built on an open agent architecture. In March it published six official skill modules covering presentation, documents, spreadsheets, design, search, and music creation.

June brought version 3.1, with a design canvas and a dynamic workflow system that decomposes a large task and dispatches dozens or hundreds of sub-agents in parallel. July brought version 3.2 and a feature that lets the agent join group chats on Slack, Lark, DingTalk, Discord, and Telegram, using conversation context to follow work requests. A companion hardware device, Skywork Note, records meetings and converts audio into notes, summaries, and drafts.

This is a product strategy built around reducing the distance between instruction and deliverable. The company is not selling access to a model; it is selling completed documents, decks, and videos. That is easier to charge for, and harder to evaluate, because the buyer must trust the output rather than a benchmark score.

Why overseas revenue is the real story

The 91% overseas share is the more interesting figure. Chinese AI companies often discuss global ambition, but few disclose such a high proportion of actual revenue from outside the domestic market. Kunlun's earlier experience as a game publisher and operator of overseas platforms gave it distribution channels and payment infrastructure that model-first startups lack.

Its consumer products follow the same logic. DramaWave is a short-drama platform; SkyReels is a creation tool for that format; Mureka makes music; Linky is an AI social product. These are categories with paying users outside China, and they generate the data and cash flow that a pure model company must raise from investors.

The company also publishes open models under the Skywork name, including multimodal reasoning and software-engineering variants. It says its agents topped the GAIA benchmark, a widely used agent evaluation. That claim comes from the company and should be treated accordingly.

In July 2026, a subsidiary called Riemann Dynamics released a "world action model" named Riemann-1.0 for embodied intelligence (具身智能), trained on first-person video, demonstration data, and robot trajectories. It is an extension of the same idea: models that act, not merely answer.

The catch with doing everything

A broader product line brings a broader set of failure modes. A short-drama platform lives on user taste and marketing spend; an office agent lives on accuracy and integration; an embodied model lives on data that is expensive to collect. Few companies can be excellent at all three, and the capital required to try is substantial.

There is also a question of what the customer is actually buying. When a vendor says an agent "completed the task," the claim usually covers a final document that a person still has to review. The economic value is real but bounded by how much verification work remains. That ratio, rather than the number of sub-agents dispatched, will decide whether these products replace software licences or remain clever demonstrations.

Honest limitations

  • Revenue figures are from the company's half-year report as summarised by China Securities Journal and the China Securities News. I did not read the original filing.
  • The 91% overseas share is a company statement. The report did not break it down by product in the sources reviewed.
  • Benchmark claims, including the GAIA result, are self-reported and not independently reproduced here.
  • Details of the agent workflow, including the number of parallel sub-agents, come from company descriptions rather than technical documentation.
  • The Riemann-1.0 model was described in a securities report; I found no independent technical evaluation.

What readers can do now

  • If you are evaluating an agent product, give it one real task with a known correct answer and compare the output with what you would have produced yourself.
  • Check where your documents are processed. Local-file handling and cloud upload have different privacy implications.
  • For investors and buyers, separate model quality from product distribution. A weaker model with a working payment channel can be a better business than a stronger one without one.

Related coverage

Cite this article

NeuroAI (脑机网), "A Chinese AI Firm Makes 91% of Revenue Overseas — and Pivots Its Products to Agents," 22 November 2026, https://neuroai.site/a/na-model-kunlun-wanwei-agent-overseas-reven

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Foundation Models” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories