Chen Danian Returns to the Arena, Entering the LLM Space
Debut Comes Close to DeepSeek’s Trillion-Parameter Flagship
By Luyu, reporting from Aofeisi Temple
Incredible! An unexpected contender has emerged among China’s top-tier domestic LLMs.
A new company has taken second place overall in the CAICT MCP Special Assessment with its first model, which has only 27B parameters.
The model ahead of it is Liang Wenfeng’s hidden trump card, kept under wraps for nearly a year: DeepSeek-V4-Pro, with as many as 1.6 trillion parameters.
The gap between the two is just 1.3 percentage points.
This dark horse is StartLux-V1.0-27B-Preview, developed by StartLux (formerly Yuandian Xinghui).
The name may sound unfamiliar, right? Hold on—the company’s founder and CEO is someone you probably know: Chen Danian.
As one of the original “godfathers” of programmers in China’s internet era, his track record is well documented:
Co-founder of Shanda Network, head of LianShang Network, and the programmer who introduced the concept of “shareware” to China at an early stage…
After ten years away from the spotlight, he has returned—this time placing his bet on local models.
That brings us to Chen Danian’s latest rare public appearance.
At the 18th-anniversary reunion of old friends from the Shanda Innovation Institute, he publicly proposed “eight non-consensus views in the AI era,” four of which emphasized local models.
Local models will completely destroy the cloud market, catch up with Claude within three years, and capture 80% of the market. The game of competing through parameter counts is OUT…
StartLux is the clearest embodiment of this philosophy.
Rather than following the industry mainstream, the company is focused on commercializing small, polished local models. It is also the world’s first genuinely local-model company in the true sense of the word.
As its first market-facing achievement, StartLux-V1.0-27B-Preview does not depend on the cloud and can run directly on a consumer-grade PC.
In other words, this local model—nearly 60 times smaller—can match the Agent capabilities of a trillion-parameter cloud flagship.
How?
27B Versus 1.6T: How a Small Model Delivered Big Results
Before revealing the answer, let’s first take a look at what StartLux was benchmarked against.
As the saying goes, no one remembers second place—unless the winner is DeepSeek.
What’s more, the gap between the two is razor-thin, making the comparison particularly worth discussing.
The results come from the authoritative China Academy of Information and Communications Technology’s (CAICT) Trusted AI Large Model Benchmark, specifically its MCP Special Assessment. The assessment defines six categories of tasks based on real-world application scenarios:
Location navigation, web search, browser automation, financial analysis, code repository management, and 3D design.
It also includes an overall evaluation focused on an Agent’s ability to coordinate multiple tools, execute complex tasks, and interact with real-world environments.
Simply put, MCP-Universe does not assess whether a model’s answers sound good. It assesses whether the task can actually be completed. This is also the most fundamental standard for evaluating an Agent’s capabilities.
The results show that StartLux-V1.0-27B-Preview achieved an overall score of 39.25%, ranking second.
It outperformed the 284B-parameter DeepSeek-V4-Flash-0731 and the 198B-parameter Step-3.7-Flash, falling behind DeepSeek-V4-Pro by just 1.3 percentage points.
Among models with the same 27B parameter scale, StartLux also scored 5.34 percentage points higher than Qwen 3.6.
Its individual scores were equally impressive: it ranked first in location navigation, financial analysis, and browser automation, while also placing near the top in the remaining subcategories.
Beyond the data, let’s look at two real-world tests.
The first involved a Microsoft stock investment spanning two years, with Claude Sonnet 4.6 as the comparison model.
Prompt: If I had invested $25,000 in Microsoft on January 9, 2023, and held it until the market close on January 8, 2025, what would the final value and total return be?
Claude’s answer was: $47,254 and 89.02%.
Video link: https://mp.weixin.qq.com/s/365CtdgGYFlEKNDICtoCfg
StartLux, meanwhile, gave the answer: $47,499.09 and 90.00%.
Video link: https://mp.weixin.qq.com/s/365CtdgGYFlEKNDICtoCfg
They may seem almost identical—but in finance, the smallest discrepancy can have enormous consequences.
A close look at the two models’ reasoning processes reveals that, because the original data was unavailable, Claude Sonnet 4.6 mistakenly judged January 8, 2025, to be a non-trading day and used the previous day’s closing price in its calculation.
Under the same circumstances, StartLux looked up the historical data, then verified market activity around the target date, confirmed the accurate closing price, and only then completed the calculation and generated a visualization.
As a result, StartLux’s conclusion was correct, verifiable, and traceable.
The second task was more straightforward: both models were asked to search for flights in a browser.
Prompt: Open Google Flights in a browser and find a one-way flight from Singapore to Beijing departing five days from now. I want the cheapest nonstop economy-class flight, but it must not arrive at Daxing Airport. Only look at the price. Close the browser when the task is complete.
StartLux-V1.0-27B-Preview completed the search in approximately 95 seconds and found an Air China ticket priced at $299.
Video link: https://mp.weixin.qq.com/s/365CtdgGYFlEKNDICtoCfg
Claude Sonnet 4.6, by contrast, went through more screenshot confirmations while selecting the date, closing pop-ups, and navigating the filter menu. At one point, it even accidentally opened the date filter panel.
More than 200 seconds later, it returned a lowest fare of $556. The price was higher, and the search took twice as long as StartLux’s.
The difference was especially clear in the action path: StartLux completed the task in just 12 steps, while Sonnet used a full 21.
Video link: https://mp.weixin.qq.com/s/365CtdgGYFlEKNDICtoCfg
This clearly shows that parameter scale is no longer the only decisive factor in Agent tasks.
The new variable is post-training.
Fewer Parameters, No Sacrifice in Capability
StartLux-V1.0-27B-Preview was not trained from scratch.
It is built on Qwen3.6-27B, yet its final test score was significantly higher than Qwen’s. The reason lies in its task data and automated post-training methods.
Put simply, StartLux trained an Agent that is better at getting things done. Its training data focuses on strengthening capabilities such as task understanding, tool selection, parameter construction, multi-step execution, state checking, and result verification.
The model must learn more than simply producing a final text response. It must also learn when to call which tool, how to adjust when a tool returns an error, and under what conditions it can declare a task complete.
Then, post-training takes the process even further.
The team independently developed an entirely new set of model iteration and optimization technologies that are multidimensional, verifiable, and scalable. One component is AI training AI (Auto Research), which allows the model to execute tasks autonomously in real tool environments and continuously adjust its strategies based on environmental feedback.
The correction and lookup process in the financial analysis example mentioned above, as well as the constraint recognition and path selection in the browser task, are the most direct demonstrations of this post-training approach.
According to the company, StartLux-V1.0-27B-Preview is also China’s first local Agent model to complete post-training using the Auto Research approach.
Of course, this does not mean that Scaling Law has become irrelevant.
Large-parameter cloud models remain the mainstream choice for now, but StartLux has sent a clear signal: they are not the only universal solution.
In Chen Danian’s words:
Scaling Law is a “passing deity.” Without it, AI would never have taken off. But it merely passes through AI’s development path; the future may not belong to it.
The industry has recognized this issue as well. The technological direction of large models is currently showing a trend toward increasingly pronounced divergence:
On one side are the faithful believers in “brute force gets results,” scaling from tens of billions to hundreds of billions and then to trillions of parameters, with training costs rising exponentially along the way;
On the other is the rapid rise of Agentic evaluation frameworks represented by Harness, with an increasing number of experts and scholars explicitly advocating that “less is more.”
Borrowing Wang Yangming’s phrase, the unity of knowledge and action is what matters—higher-quality action is the key. StartLux offers yet another example of reducing a model’s parameters without reducing its capabilities.
This also helps explain why StartLux is committed to developing local models.
Once a model runs on a PC, the cost structure of long-running tasks can shift from continuously accumulating cloud Token fees to more controllable hardware computing and electricity consumption. At the same time, the model can better understand users’ long contexts and deliver more personalized results.
StartLux is not alone. Meta, Google, and NVIDIA have also recently increased their investment in small local models.
However, most of these projects remain in the experimental stage or are geared more toward technically sophisticated users. When it comes to productizing local models, StartLux is currently the only one.
StartLux also said it will steadily move forward with training its own foundation models and explore new architectures such as diffusion language models. In short, it has chosen the right path, and its future looks promising.
StartLux Takes Up the Mantle of Local Models
Since this is a non-consensus path, its leaders need to be two kinds of people: those willing to place a bet and those capable of making it happen.
That is precisely what StartLux’s star-studded team represents: a powerful combination of entrepreneurs and scientists.
StartLux Founder Chen Danian
CEO Chen Danian has already been briefly introduced. He is a serial entrepreneur and one of China’s first-generation programmers. He rose to prominence around the same time as Zhang Xiaolong and Lei Jun, and began his entrepreneurial journey in the same era as Jack Ma and Pony Ma.
He was among the earliest people to introduce the concept of “shareware” to China. Together with his second elder brother, Chen Tianqiao, he founded Shanda Network and the Shanda Innovation Institute, one of the cradles of innovation in China’s internet industry. The mass-market product “WiFi Master Key” was also incubated under his leadership.
It can be said that he has an in-depth understanding of Chinese users and products. Local models are, in many ways, right in his comfort zone.
StartLux Co-founder Guo Quanwei
The person responsible for turning the technical vision into reality is StartLux co-founder and CTO Guo Quanwei.
Guo Quanwei holds a PhD in Computer Science and Engineering from National Yang Ming Chiao Tung University. His research covers locally deployed large models, Agentic AI, AI for Science, AI for Finance, privacy-preserving machine learning, and other fields.
Before joining StartLux, he served as Chief Algorithm Scientist at Phantom Quant, an AI science company. Earlier, he conducted research and development in data privacy, data de-identification, and privacy-preserving machine learning at Taiwan’s Industrial Technology Research Institute.
He was also the recipient of the 2024 TAAI Best Paper Award and holds invention patents related to data privacy as the primary inventor.
Another StartLux co-founder, Luo Yongxiang, is the former Managing Director of Morgan Stanley Asia.
Chen Danian understands products and users; Guo Quanwei has long focused on local models, Agents, and data privacy; and Luo Yongxiang is responsible for the market as well as fundraising and investment. This combination is highly suited to StartLux and will support its long-term development.
As for what the team ultimately wants to deliver, it is not merely a set of model weights.
In StartLux’s vision, local intelligent solutions should be deployable with a single click, much like installing Office. Users should not need to understand quantization, VRAM configuration, or inference frameworks, nor should they have to repeatedly fine-tune the system themselves.
Accordingly, the team currently plans to launch its first-generation local intelligent solutions for enterprise and individual users within this year.
Seen in this light, StartLux’s sudden emergence is far more significant than simply “adding another player to the field.”
It also represents the fact that, following the emergence of cloud-model companies such as DeepSeek and Kimi, domestic local-model companies in China are now beginning to fill the gap.
From rising stars of the intelligent era back to the first-generation programmers of the internet era, the fundamental paradigm of models is shifting—but Chinese companies have always been passing the baton forward.
Official website: https://startlux.com/