Just three months after its launch, Chen Danian's model nearly outperformed Deepseek.
48-year-old Chen Danian has started writing code again.
On August 25, 2026, during his speech at the 18th anniversary alumni gathering of the Shanda Innovation Institute, Chen Danian proposed that "the cloud-based business model is the largest-scale monopoly that humanity is facing and may occur". Five days later, StartLux, the company he founded, released its first local Agent model, StartLux-V1.0-27B-Preview.
Only three months after its establishment, StartLux has delivered a highly impressive performance report.
In the MCP special benchmark test of the Trustworthy AI Large Model Benchmark organized by the Artificial Intelligence Research Institute of China Academy of Information and Communications Technology, this 27-billion-parameter model ranked second in overall performance, only 1.3 percentage points behind DeepSeek-V4-Pro which has a parameter scale of 1.6 trillion, and outperformed DeepSeek-V4-Flash-0731 and Step-3.7-Flash whose parameter scales are far larger than itself.
Over the past two decades, Chen Danian has experienced several of the most fierce rounds of competition in China's Internet industry: from shareware to online games, from free Internet access tools to large-scale Internet products. Now, while the industry is pouring more resources into cloud models, computing power infrastructure and ultra-large parameter scales, he has bet his new company on a different path — enabling AI to run directly on users' own devices as much as possible.
This time, he predicts that within 3 years, the capability of local models will match that of Claude, and capture 80% of the market share.
Why did Chen Danian, who has been "in seclusion for ten years", make his comeback at this moment?
Starting from a free software
When Chen Danian first came into contact with the Internet, the network was far from being the infrastructure it is today.
In 1998, the Internet in China was just in its infancy. The Internet speed was very slow, while the cost was not low. Chen Danian recalled later that he took the entire family's household registration booklet to the communications administration bureau by the Suzhou Creek in Shanghai to go through the Internet access procedures. At that time, the Internet speed of two to three K per second could hardly be described as smooth. Spending two hours online every day, the monthly cost could even eat up an ordinary person's salary.
Therefore, he wrote a software named ENCounter to help users calculate the Internet access cost, and put it on the Internet for free. Soon, this software was downloaded massively, and was rated as one of the top 10 shareware in China in 1998 by PC World China.
He was 20 years old that year.
In 1999, 21-year-old Chen Danian co-founded Shanda Network with his elder brother Chen Tianqiao, with only 500,000 RMB as the initial startup capital. Two years later, Shanda obtained the Chinese distribution rights of the online game *Legend of Blood* with 300,000 US dollars, and quickly expanded the market through game operation and localized services. In 2005, Shanda took the lead to change its core games including *Legend of Blood* to the free-to-play operation mode with paid props and equipment, pushing China's online game industry into the "free-to-play era".
In 2004, Shanda was listed on NASDAQ. Chen Tianqiao and his family members topped the Hurun IT Rich List as the richest people in China with a net worth of about 8.8 billion RMB. Chen Danian was 26 years old at that time.
From a small software that solved the problem of Internet access cost to developing Internet products with huge user scales, Chen Danian's early entrepreneurship always revolved around a simple logic: first find the real demand, and then use technology to turn it into a product accessible to more people.
From Shanda Innovation Institute to WiFi Master Key
After Shanda went public, Chen Danian did not follow the traditional expansion path of Internet companies.
Around 2006, working 15 hours a day for a long time made his body severely overdrawn. One late night, he collapsed under an overpass in Pudong. After calling the emergency number, he lay on the ground, feeling that he might not make it through.
This experience made him rethink the way of entrepreneurship and work. He later concluded: "Working to the point of exhaustion for entrepreneurship is a false proposition. Excessive hard work actually does no good to entrepreneurship." Instead of making employees work overtime constantly, it is better to leave them space for learning, thinking and trying.
In 2008, Chen Danian founded Shanda Innovation Institute and served as its dean, recruiting more than 500 engineers across the country to explore new directions that had not yet shown clear commercial value. At R&D meetings, he often said: "Don't think about making money for now."
WiFi Master Key was launched in such an environment.
The project was not taken seriously at first. The first two teams left one after another, and when Zhang Fayou joined in September 2011, there were only two people left in the team. In the end, only Zhang Fayou and Chen Danian remained.
What Chen Danian saw was not just the technical problem of "connecting to WiFi". For some users, the ability to access the Internet means the ability to study, work, obtain information, and even enter a world that they could not reach before.
In 2012, Chen Danian founded Zhangmen Technology. In 2013, LianShang Network under Zhangmen Technology launched WiFi Master Key, which helped users access the Internet more conveniently by sharing and connecting to surrounding WiFi networks. By June 2016, the global users of WiFi Master Key exceeded 900 million, with 520 million monthly active users, covering 223 countries and regions.
This time, he started from a demand that was easily overlooked, and created a large-scale Internet product.
Ten years later, when Chen Danian returned to writing code, the problem he faced was no longer "how to connect more people to the Internet", but another one: how to make everyone truly own their own AI.
Founding StartLux
After 2015, Chen Danian gradually stepped down from frontline product operation, shifting his focus to investment and startup incubation. He once invested in Choudao Equity, hoping to connect entrepreneurs and investors through the Internet. At the same time, WiFi Master Key continued to expand, and he gradually handed over daily management to the team, lived in Singapore for a long time, and kept paying attention to new technologies and entrepreneurial opportunities.
In March 2025, Yusheng Science Innovation Institute was established in Shanghai, providing funding, talents, mentors and industrial resource support for AI startup projects. By its official debut in 2026, the platform had launched about 20 AI projects, covering eVTOL, robots, smart hardware, AI office, AI companionship and large model application directions. Its AI creation platform Lingzhu was first unveiled to the public, and projects including Luwu Intelligence, Photonmatrix and Yufeng Future also entered its incubation and investment ecosystem.
On May 21, 2026, Chen Danian led Yusheng Science into the public view as the chairman; on May 28, Xu Peng, former president of Lianshang Group, was promoted to CEO of Lianshang Group, taking full charge of the daily management of WiFi Master Key; immediately after that, on May 29, StartLux was officially registered and established.
Only three months after its establishment, StartLux released its first local Agent model.
Chen Danian's re-entrepreneurship is not a unique case for this generation of game entrepreneurs entering the AI era.
Chen Tianqiao, founder of Shanda, once became a representative figure in the online game era thanks to *Legend of Blood*. After leaving Shanda, he established Tianqiao Brain Science Institute (TCCI) in 2016, and has continuously increased investment in brain science and AI research since then; Zhou Yahui, founder of Kunlun Tech, started his business from web games and global game distribution, and promoted the company to shift its business from games all the way to large models, AI search, AI music, AI video and AI Native platforms.
Cai Haoyu, co-founder of miHoYo, founded Anuttacon after stepping down as chairman in 2023, starting with AI games, and then continuing to explore in the directions of LLM and Agent; Liu Dan, a veteran of Tencent Games, left Tencent in 2025 after working at Tencent Games for more than 20 years and founded Shenzhen Chuangxiang Yuedong Technology, taking AI games as his new entrepreneurial direction.
From Chen Tianqiao, Zhou Yahui, to Cai Haoyu, Liu Dan, and then to Chen Danian today, some people moved from games to brain science, some promoted their original companies to transform towards AI, some re-started their businesses and bet on AI, and others continued to cut in from the familiar entry point of games. The paths they took are different, but they are all facing the same problem: how to bring the experience, resources and judgment accumulated in the previous technology cycle into the next wave of AI.
Why local models?
When we use models such as ChatGPT and DeepSeek through online services, we usually need to send the questions to remote servers to complete the inference, and then the results are returned to users.
Local models run directly on personal computers, mobile phones and other devices, so data can stay locally, and the use does not need to continuously rely on cloud calls. Chen Danian specially emphasized that edge-side models are mostly lightweight models deployed on specific devices such as cars and smart glasses, while local models place more emphasis on running relatively complete AI capabilities on personal devices.
In the past two years, mobile phone manufacturers, enterprise service providers and startups have all been trying to migrate AI from the cloud to user devices. The difference of StartLux is that it is not satisfied with "putting the model in the computer", but further combines local operation with Agent capabilities: the model can call browsers, search engines, code repositories, financial tools, etc., to directly complete continuous tasks for users.
In Chen Danian's view, the value of local models is mainly reflected in cost, data and continuous use. Cloud AI is charged by the call volume. When AI becomes a high-frequency tool, the inference cost will continue to accumulate; after the local model is deployed, it can be used repeatedly. At the same time, data such as personal photos, work documents and family materials do not need to be fully uploaded to the cloud, which is also more direct for tasks that need to process local files and personal materials for a long time.
Chen Danian himself is an example. He is a photography enthusiast, and there are 2 terabytes of photos stored in his computer. "Can I send them to Claude to help me manage? I am absolutely unwilling. Not only am I unwilling, Claude is also unwilling. After uploading, it can't charge much Token fees, but the bandwidth and storage it occupies are massive."
Why is he so alert to cloud models?
On August 25, "Innovative Young Pioneers" released eight "non-consensus" points shared by Chen Danian at the 18th anniversary alumni gathering of Shanda Innovation Institute about the AI era. The most important one among them is his concern about cloud models.
In his view, as AI participates more deeply in work and life, cloud models no longer only master the questions raised by users, but may also include knowledge, habits, materials and even decision-making processes. The stronger the model is, the deeper the dependence between users and the platform may be.
Therefore, what he is really worried about is not the market share of a certain company, but whether users can truly own their own AI when AI capabilities and personal data are increasingly concentrated.
His answer is to bring AI back to personal devices.
Chen Danian also used the history of Sun Microsystems minicomputers and PC servers as an analogy. Once, Sun, which sold minicomputers, was one of the most concerned technology companies on Wall Street, with a market cap close to 200 billion US dollars in 2000; later, PC servers with lower cost and more flexible expansion gradually rose, and Sun was finally sold to Oracle for 7.4 billion US dollars in 2010.
In his view, AI may also experience similar changes: with the decline of computing power cost and the improvement of model efficiency, expensive centralized computing will not necessarily occupy the dominant position forever.
He even called the ultra-large parameter models that the industry is competing to pursue now "passing immortals".
"Without it, AI cannot get started, so we should respect it and be grateful to it. But it is just passing by on this road, and the future of AI definitely does not belong to it."
He also proposed that the power consumption of the human brain when solving complex problems is about 20 watts, while the computing power consumption required for large model operation may reach millions of watts.
"Why do machines burn so much more electricity than humans, and are not necessarily smarter than humans?"
In Chen Danian's view, this means that there is still huge optimization space for existing algorithms and computing methods.
27B Model,
How to complete complex tasks?
The first model released by StartLux is Chen Danian's technical response to the local AI route.
StartLux-V1.0-27B-Preview is based on Qwen3.6-27B as the base, and after post-training for Agent capabilities, it supports running on consumer-grade personal computers. In the trustworthy AI large model benchmark test organized by the Artificial Intelligence Research Institute of China Academy of Information and Communications Technology, it ranked second overall in six types of tasks including location navigation, web search, browser automation and financial analysis, with a score only 1.3 percentage points lower than DeepSeek-V4-Pro with 1.6 trillion parameters.
StartLux also announced two same-task tests against Claude Sonnet 4.6.
In the task of calculating Microsoft stock, after finding that the target date data was missing, StartLux re-queried, and finally gave the result of 47499.09 US dollars with 90.00% accuracy; Claude gave 47254 US dollars with 89.02% accuracy. In the flight ticket search task, facing the requirement of "the cheapest one-way direct economy class from Singapore to Beijing on the specified date, excluding Daxing Airport", StartLux performed 12 steps of operations, took 1 minute and 31 seconds, and found a compliant flight priced at 299 US dollars; Claude performed 21 steps of operations, took 3 minutes and 42 seconds, and gave the lowest price of 556 US dollars.
What is really worth paying attention to in the two tests is not just the final answer, but whether the model can judge: the result returned by the tool does not mean the task has been completed. StartLux will continue to check the data integrity and condition compliance, and re-execute after finding problems.
To a large extent, this comes from post-training. StartLux adopts the method of "AI trains AI", allowing the model to execute tasks in the tool environment, and then adjust the training strategy according to feedback. The team also studies the cross-layer activation rule in the mixed linear attention large model, explores model compression, quantization and inference efficiency optimization, to reduce the computing power and resources required for local operation while retaining capabilities as much as possible.
At present, StartLux-V1.0-27B-Preview is still in the preview stage, and the model filing and productization work are underway. The team plans to launch the first generation of local intelligent solutions within the year.
From shareware to online games, from WiFi Master Key to local AI, what Chen Danian has always focused on is how technology can cross the usage threshold and enter more people's daily lives. In the early stage, the threshold was the cost of Internet access; later, it was network connection; now, it may be the usage cost, data control right and deployment mode of AI.
Local models do not mean that cloud models will disappear. A more realistic direction is that the two form a division of labor: complex and heavy tasks continue to rely on the cloud, while scenarios involving personal files, private data, continuous operation and offline use are undertaken by local models.
Conclusion
In his speech on August 25, Chen Danian said that AGI will not be realized by relying on a supreme "god", but by 8 billion people around the world. Each model is not smart enough, but because of their ever-changing forms and distinct personalities, they can make continuous breakthroughs.
He said: "The real AGI moment should be when it finds the path that everyone else thinks is wrong, but it thinks is right."
This sentence can also be used to understand this entrepreneurship of his.
Large cloud models are still evolving rapidly, and local models are far from proving that they can become the mainstream. But Chen Danian has already begun to try another answer: AI does not necessarily have to exist in huge cloud infrastructure, and it can also become a long-term capability in personal devices.
StartLux's goal is very direct: to realize AGI and ensure that it belongs to the people.
This article is