Stable AI
Category: Foundation Models / LLMs
Chinese startup (稳准智能/Wenzhun Intelligence) building LimiX, a general-purpose large data model (LDM) for structured/tabular data, co-developed with Tsinghua University's Prof. Cui Peng team. Stable AI was founded in 2024. The company is led by Zhang Tianyang (张天扬). Based in Xiong'an New Area / Beijing, China. Team size: 11-50. Latest round: Series A. Key investors include Xiong'an Science & Technology Innovation Angel Fund Phase I, Oriental Fortune Capital (东方富海), MSA Capital (和玉资本), Guoke Investment (国科投资), Beijing Guoke Zhengdao Investment Center, Beijing Guoke Ruiqi Innovation & Entrepreneurship Investment Fund.
- Founded
- 2024
- Headquarters
- Xiong'an New Area / Beijing, China
- Team size
- 11-50
Value proposition
A general-purpose Large Data Model (LDM) for structured/tabular data that performs classification, regression, missing-value imputation, causal inference, and OOD generalization out-of-the-box with zero training, delivering explainable, causally-grounded decision intelligence for industrial and enterprise scenarios.
Products and solutions
LimiX (极数) general structured-data foundation model, LimiX-2 (400M-parameter, released Sep 2026), LimiX-2M (2M-parameter, open-sourced Nov 2025), YiZhun (易准) electricity-trading data foundation model (with Zhongwang Union), industry vertical solutions for manufacturing, energy, transportation, telecom, petrochemical, metallurgy.
Unique value
First unified foundation model for structured data (LDM) integrating structural causal inference with pretrained large-model technology — one model covers 10+ task types across industries with causal interpretability, no retraining, and SOTA performance on tabular benchmarks (TabArena, BCCO, TALENT).
Target customer
Industrial enterprises (manufacturing, energy/power, petrochemical, metallurgy, transportation, telecom), plus medical and financial sectors; developers via open APIs; researchers via open-source models.
Industries served
Manufacturing, Energy & Power, Petrochemical, Metallurgy, Transportation, Telecommunications, Medical, Finance
Technology advantage
Contextual Mechanism Network (CMN) paradigm; Context-Conditional Masked Modeling pretraining; structural causal inference / causal stable learning (pioneered by Prof. Cui Peng); zero-training out-of-the-box generalization; causal interpretability and counterfactual-falsifiable decision logic; edge-deployable high-efficiency inference.
How they differentiate
Unlike LLMs that struggle with structured data, LimiX is purpose-built for tabular/time-series data with causal inference baked in, offering one-for-all task coverage, zero-training deployment, and transparent causal decision logic — differentiating from dedicated per-scenario models and from text-centric LLMs.
Main competitors
TabPFN (Prior Labs), TabICL, Mitra, Orion, Google, SAP, and Amazon tabular foundation models
Key partnerships
Tsinghua University (Prof. Cui Peng's team, co-developer), Zhongwang Union (中网联合, YiZhun electricity-trading model), Richinfo (彩讯股份, telecom vertical model), Xiong'an New Area government
Notable customers
Industrial enterprises across manufacturing, energy/power, petrochemical, metallurgy, and telecom verticals in China, Henan Province industrial enterprises, Xiong'an New Area initiatives
Major milestones
Founded Oct 2024 in Xiong'an New Area, Angel round Feb 2025, LimiX open-sourced Aug 2025 (first general structured-data foundation model), Pre-A round Sep 2025, YiZhun electricity-trading model Oct 2025, LimiX-2M open-sourced Nov 2025, A round Jun 2026, LimiX-2 (400M) released Sep 2026 topping TabArena/BCCO/TALENT benchmarks.
Growth metrics
LimiX validated on 800+ datasets; 10+ task categories; LimiX-2 tops TabArena (Elo 1935), BCCO (Elo 1432), TALENT (Elo 1506).
Market positioning
Claims first place on TabArena (Elo 1935), BCCO (Elo 1432), and TALENT (Elo 1506) tabular benchmarks, ahead of Google, SAP, and Amazon tabular models; positioned as China's first and globally leading general structured-data foundation model (LDM).
Geographic focus
China (Xiong'an New Area, Beijing); expanding industrial verticals domestically
Patents and IP
Registered trademarks including "STABLE AI" and "稳准智能" (Class 42); multiple registered trademarks per Baidu Baike.
About Zhang Tianyang (张天扬)
Co-Founder & CEO of Stable AI (稳准智能); Tsinghua University education.
Latest news about Stable AI
More Foundation Models / LLMs companies
Official website: https://www.stable-ai.ai/