15:I[981,["658","static/chunks/658-a698693f56d0116a.js","183","static/chunks/183-6beeeb801c274aaa.js","474","static/chunks/app/%5Btype%5D/layout-631c4730b83c8223.js"],"Posts"] 16:T6b9,I’m excited to announce Scaling Mode for TabPFN-2.5, our tabular foundation model (Large Tabular Model). This removes any fixed upper limits on dataset sizes and extends TabPFN to datasets with millions of rows.

What is Scaling Mode? A new pipeline around TabPFN-2.5 designed for large-N workloads. It removes the fixed row limit. The system works with large training sets, constrained only by your compute and memory.

Performance: We benchmarked Scaling Mode on datasets ranging from 1M to 10M rows, comparing against CatBoost, XGBoost, and LightGBM. Key findings:

- Scaling Mode enables TabPFN-2.5 to continue improving performance with more data

- Scales dramatically better than TabPFN-2.5 with 50K subsampling

- No evidence of the performance gap to gradient boosting shrinking as we scale up

- Performance continues to improve strongly with more data on the largest tested datasets

The scaling trajectory:

- TabPFN v2 (Jan 2025): 10K rows

- TabPFN-2.5 (Nov 2025): 50K rows

- Scaling Mode (today): 10M+ rows tested

Current limitations: Scaling Mode is designed for companies for now. If you’re working with data at a large scale and want to test it, access is currently only by request to ensure we can support early users properly.

Full blog post: *https://priorlabs.ai/technical-reports/large-data-model**

Request access: *https://priorlabs.ai/tabpfn/large-data** Would love to hear feedback from anyone working with large tabular datasets.

Show HN: TabPFN Scaling Mode – Tabular Foundation Model on Rows

Toward a Working Definition of Paperclip-Punk

LLMs as a new alien subset of all possible embodied intelligences

Devtools Just Became AI Infrastructure

Show HN: Informiton – Intelligent Healing Signals

Show HN: GPTShirt – Custom t-shirts powered by Nano Banana Pro

Bio-essential sugars in samples from asteroid Bennu

Remodeling My Home Proved One Thing: We Need Robots to Build Houses

New Elf-Stats Malware Campaign on NPM

33 days, new WR in speedrunning a fan translation of a long-lost game

OpenAI Agrees to Acquire Neptune to Improve AI Model Training

Power Outage: On the Thermodynamics of History

The Metaverse Only Has 900 Users [video]

Sunflower Basic

The Triumph of Logical English

US pressures Mexico for violating the 1944 Water Treaty

Ghostty Is Now Non-Profit

Formally Verifying Advent of Code Using Dijkstra's Program Construction

Micron to exit consumer memory business amid global supply shortage

Private employers shed 32,000 jobs in November

A central hub for LLM API config info: model-api.info

The Quiet Tensions of Specialized Roles

Authors retract Nature paper projecting high costs of climate change

Towards a Language for Optimization

Aerosol induction-removal system for mitigating airborne disease transmission

Antiidle – A Free, Open-Source Anti-Idle Tool

From Moderation to Mediation: Can LLMs Serve as Mediators in Online Flame Wars?

Waymo Now Autonomous in Dallas

Kilo Deploy: Ship Apps Directly from Kilo

TSA's New $45 Fee at U.S. Airports Unfairly Punishes Families in the Fine Print