Stable AI and Tsinghua release LimiX-2, claiming first place on three tabular foundation-model benchmarks
The AMW Read
Claimed SOTA on three structured-data benchmarks plus an explicit scaling-law thesis at 400M meaningfully updates the foundation-model player map for the tabular modality.
Stable AI and Tsinghua release LimiX-2, claiming first place on three tabular foundation-model benchmarks
Stable AI (稳准智能), working with Tsinghua University computer science professor Cui Peng, released LimiX-2 on September 16, 2026 — a structured-data foundation model scaled to 400 million parameters. The company reports overall Elo scores of 1935 on TabArena, 1432 on BCCO, and 1506 on TALENT, placing first on each leaderboard ahead of peer models from Google, SAP, and Amazon. The upgrade centers on Contextual Mechanism Networks for joint variable-dependency modeling, a stronger automated synthetic-data engine for pretraining, and further parameter scaling after the team previously argued that structured-data foundation models follow a scaling law. Weights and code are posted on Hugging Face, ModelScope, and GitHub, with a technical report on arXiv (2609.17488).
Tabular foundation models sit outside the text-and-multimodal spotlight, yet they compete for the same enterprise prediction work that still runs on classical machine learning. A Chinese lab posting top Elo against Google, SAP, and Amazon on shared structured-data benchmarks sharpens that race and reinforces the claim that this modality has its own scaling curve, not just prompt-tuned LLMs wrapped around tables. Coverage of Stable AI is just emerging — one item in the last 90 days versus none in the prior window, per the AI Market Watch index (name-matched over pipeline-ingested sources only).
Builders should treat LimiX-2 as a check against TabPFN-class baselines for classification, regression, imputation, and causal-skeleton tasks before defaulting to per-dataset specialist models. Investors watching enterprise AI should note the go-to-market already cited — process optimization, equipment fault prediction, power and energy forecasting, materials science — and the stated path from prediction toward causal decision support, while treating leaderboard Elo as necessary but not sufficient evidence of production lift.
