Lexsi Labs unveils Orion-MSP, a new state-of-the-art Tabular Foundation Model transforming how Enterprises harness AI for Structured Data

Poised to deliver enterprise-scale accuracy for organizations to make tabular predictions at a level that existing LLMs simply can’t match.

Mumbai, Paris, London — 12th November 2025 — Lexsi Labs, by Aurionpro (formerly AryaXAI Alignment Labs), the research lab dedicated to frontier research in AI Alignment and Interpretability, today announced two major developments accelerating enterprise-scale capabilities for structured/tabular data into mainstream AI.

1. Launch of Orion-MSP: Lexsi Labs launches new state-of-the-art (SOTA) foundation models to deliver enterprise-grade accuracy and scalable deployment for any tabular predictive task. 
2. Release of TabTune Open-Source Library:  A first-of-its-kind open-source tool that standardizes fine-tuning and inferencing for leading tabular foundation models, including Orion-MSP. 

The launch of Orion-MSP and TabTune marks a significant step in Lexsi Labs’ mission to build “Safe SuperIntelligence” systems that are transparent, aligned, and trustworthy for enterprise use. Enterprises such as banks, insurers, and healthcare organizations generate large volumes of structured, table-based data from transactions, customer records, insurance claims, health records, sensor telemetry, and ledgers. Lexsi Labs’ latest models allow organizations to predict any task in zero-shot, on such enterprise data, using a single model through an API or improve it further through fine-tuning to achieve the best performance in just a few lines of code. This transforms the current data science from multi-pipeline processes into a zero-shot or few-step deployment. 

“All current AI research is primarily aiming to predict the world in ‘zero-shot’. We are seeing this achieved by modality-specific SOTA models like LLMs for text, LRMs for reasoning tasks, and LVMs for vision tasks. We want to make this possible as well for tabular predictive tasks with Orion-MSP class of models. Orion-MSP is the state-of-the-art model, with top mean-rank accuracy across various benchmarks compared to other TFMs, including classic Machine Learning models,” says Vinay Kumar, Founder & CEO of Arya.ai and Founder of Lexsi Labs. “We are also launching TabTune, which is the perfect tool for any practitioner or enterprise to infer or fine-tune TFMs. With advanced components like fairness and conformity index, this is purpose-built for enterprises. We are releasing these classes of models under the MIT license, making both the models and the tool fully usable without constraints.”

About Orion-MSP: The inaugural Orion-MSP model deploys multi-scale processing, block-sparse row-wise attention, and Perceiver-style latent memory—delivering efficient modeling of wide, heterogeneous tables while reducing computational complexity. Early results show competitive performance on large feature sets and real-world datasets, outperforming the current TFMs on several global benchmarks. 

Quick access: Orion-MSP is available on Hugging Face from today, with full research details published on arXiv. 

About TabTune:  TabTune is a novel open-source tool that offers data science teams a single, easy-to-use tool for adapting, evaluating, and deploying powerful tabular foundational models. With TabTune, we introduce multiple fine-tuning methods for TFMs such as 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 (LTM).

Quick access:  It is available on GitHub; the technical paper is on arXiv.