Buyers and investors following China's AI Plus rollout get a named alternative to chatbot-style models for factory and financial data: LimiX, now in a second version, per its developer.
Language models are "very difficult to use" for industrial work, Tsinghua University associate professor Cui Peng told the Hong Kong International Computer Conference on Monday, November 9, according to the South China Morning Post, which Alibaba owns. Cui said models trained mainly on internet text struggle to capture patterns in structured numerical data, such as factory sensor readings and financial records, and that specialist systems for that data could become a key part of the next phase of AI in China.
His team’s answer is LimiX, built with Beijing-based Stable AI (稳准智能), which the SCMP calls Wenzhun Intelligence. Cui described it as a "large data model" trained only on synthetic data. He said it can identify causal patterns and handle industrial and scientific tasks, including controlled nuclear fusion, and that it cuts the cost and time of deploying AI systems by about 90 percent while improving accuracy by more than 15 percent on average. Those figures are Cui’s; the SCMP text offers no outside test.
The remarks arrive as Beijing pushes its AI Plus program into manufacturing, logistics and healthcare. LimiX models tables of numbers and is trained on generated data, with no internet text. That makes it a different product from the chatbots that dominate the global AI market, and it is aimed at the kind of factory data the State Council plan wants AI to reach.
LimiX treats tables of numbers as the thing to model
36Kr reported that the team open-sourced LimiX on August 29, 2025, calling it the first general-purpose large model for structured data. It is trained entirely on data generated from structural causal graphs, with no real-world data. The model handles classification, regression, filling in missing values, data generation and causal inference without task-specific training, and it can be fine-tuned for further gains. The 36Kr article cites developer-reported results: a 15 percent gain in equipment-failure prediction accuracy at a steel company over the previous specialized model, and a fivefold gain in regulation efficiency at a materials company.
The team’s technical paper describes two versions, LimiX-16M and LimiX-2M, which treat a table as a joint distribution over variables and missing values. The authors tested them on 11 large structured-data benchmarks and released all models under the Apache 2.0 license.
A larger successor followed. QbitAI reported on September 16, 2026, that Stable AI (稳准智能) and Tsinghua released LimiX-2 at 400 million parameters. It ranked first on three structured-data benchmarks, with Elo scores of 1,935 on TabArena, 1,432 on BCCO and 1,506 on TALENT. QbitAI said the earlier LimiX had been validated in more than 800 structured-data scenarios and had strategic partnerships with over 60 customers. These are the developer’s claims as relayed by QbitAI.
Local governments and Huawei describe an industry-wide model of adoption
Cui was one of several speakers who described China’s approach as system-wide. Nie Jun Yu, Huawei’s global chief architect of digital government, said at the same conference that China has moved from isolated AI applications to industry-wide systems, often coordinated by local governments. In provinces such as Zhejiang and Guangdong, he said, local authorities coordinate AI adoption across whole industries, so individual companies do not build their own. He also said China puts more emphasis than the United States and Europe on making AI widely accessible through infrastructure, and described AI as "a new digital public good."
The national plan sets the targets. The State Council’s opinions on implementing the "AI Plus" action (《国务院关于深入实施“人工智能+”行动的意见》), published on August 26, 2025, aim for AI penetration of more than 70 percent in new-generation smart terminals and agents by 2027 and more than 90 percent by 2030. The document also calls for AI across industrial design, pilot testing, production, service and operations, and for smarter coordination of industrial supply chains.
Deployment is visible in logistics. China Daily, the state-run newspaper, reported on November 9 that a China Federation of Logistics and Purchasing report put AI adoption in logistics and supply chains above 37 percent, with transport at 78 percent and smart warehousing at 47.27 percent. JD Logistics said AI-recommended delivery routes have an adoption rate above 90.2 percent.
Compute and a new Hong Kong plan sit behind the next phase
Other speakers pointed beyond industrial data. Tencent’s chief AI scientist, Vinces Yao Shunyu, said on November 7 that the company plans to develop "autonomous science labs" able to run complex experiments, including protein research, the SCMP reported.
The conference took place in Hong Kong, which is positioning itself as a place to turn such advances into productivity. Sun Dong, the city’s Secretary for Innovation, Technology and Industry, said at the opening that Hong Kong is expanding AI computing infrastructure and strengthening safeguards to speed adoption across industries. Chief Executive John Lee Ka-chiu said earlier this year that the city is betting on a 36-fold increase in compute capacity by 2032, and Hong Kong is preparing its first five-year plan aligned with Beijing’s, which prioritizes technological innovation and AI integration.
The test for LimiX is outside validation. A named manufacturer or utility publishing before-and-after results, or an independent benchmark of the 90 percent cost claim, would show whether the approach scales beyond the developer’s own pilots.