LimiX
Summary
LimiX is an open tabular foundation-model family from Stable AI and Tsinghua University collaborators. It learns conditional inference over static tables from hierarchical structural-causal synthetic tasks, then performs classification, regression, and related completion tasks from in-context rows without per-dataset gradient training.
Model Family
| Model | Parameters | Main role | Evidence source |
|---|---|---|---|
| LimiX-16M | 16.52M | Highest-accuracy general LimiX checkpoint in the original report; supports classification, regression, imputation, generation, and retrieval-based inference experiments. | LimiX technical report |
| LimiX-2M | 1.92M | Compact classification/regression checkpoint using RaBEL numeric tokenization and sample-first, readout-aligned bidirectional attention. | LimiX-2M |
Why The Compact Model Is Strong
LimiX-2M combines four forms of specialization:
- Hierarchical-SCM pretraining compiles a static-tabular inference prior into the weights.
- Context-Conditional Masked Modeling turns context rows into per-dataset non-parametric memory.
- RaBEL expands each numeric scalar into localized radial-basis responses, reducing shallow low-rank value bottlenecks.
- Sample-attention FFN feature-attention routing and attention pooling make cross-sample statistics directly useful to the readout.
The result is parameter-efficient, not learning-effort-free: pretraining data and compute, context processing, retrieval passes, and ensemble pipelines are outside the 1.92M parameter count.
Official Artifacts
- Project: https://www.limix.ai/
- Code: https://github.com/limix-ldm-ai/LimiX
- LimiX-16M checkpoint: https://huggingface.co/stable-ai/LimiX-16M
- LimiX-2M checkpoint: https://huggingface.co/stable-ai/LimiX-2M
- LimiX-2M ICML 2026 page: https://icml.cc/virtual/2026/poster/63251
The repository code is Apache-2.0. The project documentation states that model weights have separate terms, including authorization requirements for commercial use, so code and weight licensing SHOULD NOT be conflated.
Role In The Wiki
LimiX is a static-tabular analogue for three time-series research questions:
- how synthetic structural priors can train a reusable in-context inference procedure;
- how scalar numeric tokenization can waste or expose hidden-state capacity;
- how compact models can beat larger models by aligning inductive bias and readout routing with the task.
It is not a time-series model. Its sample axis contains table rows rather than ordered observations, and the released evidence does not establish temporal next-state prediction, event-stream processing, or action-conditioned world-model rollouts.
Evidence
- LimiX: Unleashing Structured-Data Modeling Capability for Generalist Intelligence
- LimiX-2M: Mitigating Low-Rank Collapse and Attention Bottlenecks in Tabular Foundation Models
Relation To Foundation TSFM Agenda
Use the source-level agenda mappings in LimiX and LimiX-2M. At the entity level, LimiX is evidence for learned tabular inference priors, localized scalar embeddings, and architecture-level parameter efficiency. It is adjacent rather than direct evidence for time-series foundation models.