Lexsi Labs, the research lab dedicated to frontier research around AI Alignment and Interpretability within Arya.ai—an Aurionpro company, announced three major developments: the release of its first tabular foundational model Orion-MSP, the open-source tool TabTune, and its rebranding and expansion to London. These steps mark a pivotal moment in bringing enterprise-scale capabilities for structured data into the AI mainstream.
With enterprises generating vast volumes of table-based data from transactions, claims, sensors and ledgers, Lexsi Labs’ new offerings aim to bridge the gap between unstructured AI models and the structured reality of business data enabling data scientists and engineers to deploy powerful, interpretable models at scale.
‘Orion-MSP’, the first model in the ‘Orion’ family of tabular foundation models (TFMs) that ranks the best compared to the current TFMs for multiple tabular benchmarks (TALENT, OML, TabZilla). It introduces a new state-of-the-art (SOTA) approach combining multi-scale processing, block-sparse row-wise attention, and a Perceiver-style latent memory to model both local and global dependencies efficiently across wide, heterogeneous tables. This approach is proven to be efficient and highly scalable across multiple types of tabular datasets. ‘Orion-MSP’ offers unique advantages by operating on complex, real-world datasets with large features in Zero-shot, compared to current TFMs such as TabPFN, TabICL, TabDPT, Mitra etc.
Lexsi Labs also released TabTune, a one-of-a-kind open-source tool that standardizes preprocessing, adaptation, and evaluation for leading tabular foundation models through a single scikit-learn–style API. It supports zero-shot inference, meta-learning, supervised fine-tuning (SFT), and parameter-efficient fine-tuning (PEFT/LoRA), with built-in calibration (ECE, MCE, Brier) and fairness diagnostics, plus benchmarking harnesses across major datasets like TALENT, OpenML-CC18, and TabZilla. This will not only allow practitioners, data scientists etc. to use TFMs in enterprise scale but also allow LLMs to predict more accurately in Zero-Shot more reliably.
As part of its evolution, the research arm has rebranded from AryaXAI Alignment Labs to Lexsi Labs – short for Le eXplainable Safe Superintelligence Labs, reflecting its mission to build interpretable, safe, and aligned AI systems. With research hubs in Mumbai, Paris, and now London, Lexsi Labs deepens its global commitment to developing AI systems that reason deeply, learn responsibly, and remain aligned with human values, even as they scale.
Mr. Vinay Kumar, Founder & CEO, Arya.ai and Founder of Lexsi Lab
Tabular datasets form a very fundamental way of capturing information. Enterprises live on tables like transactions, claims, ledgers, sensors, yet most of the current LLMs are built for unstructured text and images. Alternatively, the performance of current traditional ML involves carefully creating task specific pipelines from data preprocessing to hyper parameter tuning to deliver best performance, which is a heavy AI engineering task,” said Mr. Vinay Kumar, Founder & CEO, Arya.ai and Founder of Lexsi Lab
“All the current AI research is primarily aiming to making predict the world in ‘Zero-shot. We want to make it possible with Orion models for tabular predictive tasks. Orion-MSP consistently maintained a top mean-rank in accuracy on various benchmarks compared to other TFMs (like TabPFN, TabICL, TabDPT, Mitra) and Classic ML (GBMs). It delivers enterprise scale accuracy on structured enterprise data enabling organizations to make predictions and decisions with a level of precision, and consistency, that existing LLMs simply can’t match and generalize across problem statements which classic ML lacks.”
He added, “‘TabTune’, is a novel opensource tool offers data science teams a single, easy-to-use tool for adapting, evaluating, and deploying powerful tabular foundational models. With ‘TabTune’, we introduced multiple fine-tuning methods for these TFMs like Meta-learning, SFT and, PEFT, making it heavily utility focused. This will allow teams to rapidly compare inference-only vs. tuned strategies, weigh memory/latency trade-offs, and audit calibration and fairness without rewriting pipelines for every large tabular model (LTMs). It’s the first and one of a kind tool, like a control plane for tabular foundation models (TFMs). We hope to democratize the use of TFMs as much as scikit-learn did for machine learning and provide superior predictive capabilities to LLMs.”
Availability
- Orion-MSP is available on HuggingFace from today, with full research details published on arXiv.
- TabTune is released in open source, and comes with guided examples, model-aware preprocessing, a unified tuning controller, and leaderboard utilities for reproducible comparisons. It is accessible in github and the technical paper at arXiv.
The launch of Orion and TabTune marks a significant step in Lexsi Labs mission to build “Safe Intelligence” AI systems that are transparent, aligned, and trustworthy for enterprise use. With R&D hubs in Mumbai, Paris, and London, Lexsi Labs brings together top AI researchers to bridge the gap between AI alignment research and practical enterprise deployment. With these open releases, Lexsi Labs aims to advance community-driven research and adoption of interpretable, scalable AI on structured data globally.
If you have an interesting Article / Report/case study to share, please get in touch with us at editors@roymediative.com roy@roymediative.com, 9811346846/9625243429.
