Chen Danian returns to AI with StartLux: its 27B local model nears DeepSeek's trillion-parameter flagship

A surprise has surfaced at the top of China's foundation-model race: a new company's debut model, with only 27B parameters, took second place overall in the CAICT MCP-specific evaluation. The only entry ahead of it was DeepSeek-V4-Pro, the roughly 1.6-trillion-parameter flagship, separated by a razor-thin 1.3 percentage points. The dark horse is StartLux-V1.0-27B-Preview, from startup StartLux, whose founder and CEO is a familiar name returning after nearly a decade away: Chen Danian.

According to a QbitAI report, Chen Danian is a co-founder of Shanda, the man behind Lianshang Network and the national hit app WiFi Master Key, and one of China's first-generation programmers. This time he is betting on local models: instead of following the industry's parameter race, StartLux focuses on commercializing small, local models, describing itself as the world's first true local-model company. StartLux-V1.0-27B-Preview runs without cloud dependency and can execute directly on consumer-grade PCs.

The benchmark results come from the MCP-specific track of the CAICT trusted-AI large-model evaluation, which tests six real-world task categories: location navigation, web search, browser automation, financial analysis, code repository management, and 3D design. StartLux-V1.0-27B-Preview posted an overall score of 39.25% for second place, beating DeepSeek-V4-Flash-0731 (284B) and Step-3.7-Flash (198B), and outpacing Qwen 3.6 by 5.34 percentage points at the same 27B scale. It also ranked first in location navigation, financial analysis, and browser automation.

According to the company, the model is not trained from scratch: it starts from a Qwen3.6-27B base and differentiates itself through task data and automated post-training. Training emphasizes task understanding, tool selection, parameter construction, multi-step execution, state checking, and result verification, teaching the model when to call which tool, how to recover from tool errors, and when a task is truly done. The post-training pipeline uses a proprietary, verifiable, scalable iteration method built on an "AI trains AI" Auto Research approach, in which the model executes tasks in real tool environments and adjusts based on environmental feedback; it is described as the first domestic local agent model post-trained this way.

The article also details two head-to-head demos against Claude Sonnet 4.6. On a two-year Microsoft stock investment question, Claude mistook January 8, 2025 for a non-trading day and answered $47,254 with an 89.02% return, while StartLux re-checked the data, verified prices around the target date, and produced a verifiable $47,499.09 and 90.00%. On a Google Flights task to find the cheapest one-way flight from Singapore to Beijing, StartLux found a $299 Air China fare in about 95 seconds across 12 steps, while Claude took over 200 seconds and 21 steps and returned a $556 minimum. That is enough to show that on agentic tasks, parameter scale is no longer the only decisive variable.

StartLux is led by an entrepreneur-scientist pairing: co-founder and CTO Guo Quanwei holds a computer science PhD from National Yang Ming Chiao Tung University, previously served as chief algorithm scientist at AI science company Huanliang Technology, and won a 2024 TAAI best paper award; co-founder Luo Yongxiang is a former Morgan Stanley Asia managing director handling market and fundraising duties alongside Chen Danian's product and user expertise.

The company's vision is for local intelligence to deploy as easily as installing Office, with no need for users to understand quantization, VRAM configuration, or inference frameworks. It plans to release its first-generation local intelligence solution for enterprises and individuals within the year, while steadily advancing self-trained base models and exploring architectures such as diffusion-style language models.

With cloud-model players like DeepSeek and Kimi already established, StartLux signals that Chinese local models are beginning to fill the gap, offering a fresh data point for the "fewer parameters, same capability" thesis. Notably, Meta, Google, and NVIDIA have also been adding weight to small local models lately, though most efforts remain experimental, and few players are as focused on productizing local models. The questions to watch: whether the first-generation local intelligence solution ships as promised this year, and whether a 27B model can translate benchmark scores into real commercial deployments.