--- title: Tabular LLMs: An Introduction to the Foundation Models... description: Tabular foundation models — pretrained transformers that predict on any spreadsheet zero-shot via in-context learning — now outperform fully tuned... canonical: https://daily.dev/posts/tabular-llms-an-introduction-to-the-foundation-models-that-predict-your-spreadsheet-dkoxe4tky twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Tabular LLMs: An Introduction to the Foundation Models That Predict Your Spreadsheet | daily.dev og:description: Tabular foundation models — pretrained transformers that predict on any spreadsheet zero-shot via in-context learning — now outperform fully tuned... og:url: https://daily.dev/posts/tabular-llms-an-introduction-to-the-foundation-models-that-predict-your-spreadsheet-dkoxe4tky og:image: https://api.daily.dev/og/posts/dkoXE4tky.png og:image:alt: Tabular LLMs: An Introduction to the Foundation Models That Predict Your Spreadsheet og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Tabular LLMs: An Introduction to the Foundation Models That Predict Your Spreadsheet **[Towards Data Science](https://daily.dev/sources/tds)** · 23 min read · 0 upvotes · 0 comments ## Summary Tabular foundation models — pretrained transformers that predict on any spreadsheet zero-shot via in-context learning — now outperform fully tuned gradient-boosted trees on the TabArena benchmark. The author independently reproduced TabICLv2 (28M parameters, BSD-3 open weights, synthetic-only pretraining) on 51 datasets for $2 of GPU time, landing within bootstrap noise of the official score. Key findings: TabICLv2 sits 158 Elo above tuned+ensembled LightGBM; it fits in seconds with no hyperparameter search; it beats a prompted Claude Opus on 43 of 49 datasets; but it loses to GBDTs on wide tables (>100 features) and high-cardinality categoricals. A validation-based router between two foundation models recovers ~half the oracle headroom, while routing between a foundation model and a GBDT actually hurts accuracy. Contamination concerns are flagged for models trained on real OpenML data, while synthetic-only models like TabICLv2 offer the cleanest evidential position. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### Does TabICLv2 outperform tuned gradient-boosted trees like XGBoost or LightGBM on tabular data? Yes, on the TabArena benchmark TabICLv2 sits 158 Elo above tuned-and-ensembled LightGBM (1432) despite requiring no hyperparameter tuning, and it wins 40 of 51 datasets against the best tuned-and-ensembled GBDT. It loses mainly on tables with more than 100 features or high-cardinality categorical columns, where trees remain the better choice. _Anyone choosing between GBDTs and foundation models for tabular data can track these tradeoffs on daily.dev._ ### How many parameters does TabICLv2 have and what license is it released under? TabICLv2 has 27,552,258 parameters for its classifier checkpoint and 28,544,991 for its regressor, roughly a quarter the size of the t0-alpha time-series model. It is released as open weights on HuggingFace under a BSD-3 license with no gating or click-through requirement, and installs via pip install tabicl. _Developers evaluating open-weight tabular models can follow licensing and size details like this on daily.dev._ ### Are tabular foundation models faster or cheaper to run than tuned gradient-boosted trees? TabICLv2's median fit time is about 4 seconds per 1,000 rows, roughly a hundred times shorter than tuned GBDT protocols, since fitting is a single forward pass with no hyperparameter search. Its median predict time, 0.38 seconds per 1,000 rows, beats tuned-and-ensembled LightGBM's 2.64 seconds, though it runs on GPU while default CatBoost remains the cheapest option on CPU at 0.08 s/1K. _Teams weighing serving costs between foundation models and trees can compare benchmarks like this on daily.dev._ ## Similar posts on daily.dev - [The Magic of In-Context Learning \(ICL\): When Your Model Already Knows Your Data](https://daily.dev/posts/the-magic-of-in-context-learning-icl-when-your-model-already-knows-your-data-jwwujzzfn) · R-bloggers · 0 upvotes · 0 comments - [One Model to Rule Them All? SAP-RPT-1 and the Future of Tabular Foundation Models](https://daily.dev/posts/one-model-to-rule-them-all-sap-rpt-1-and-the-future-of-tabular-foundation-models-9r5j35o1p) · Towards Data Science · 0 upvotes · 0 comments - [NVIDIA Kumo Tabular Sets a New Accuracy-Efficiency Frontier for Tabular Prediction](https://daily.dev/posts/nvidia-kumo-tabular-sets-a-new-accuracy-efficiency-frontier-for-tabular-prediction-avcjvvwsc) · Hugging Face · 0 upvotes · 0 comments --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning) [View this post on daily.dev](https://daily.dev/posts/tabular-llms-an-introduction-to-the-foundation-models-that-predict-your-spreadsheet-dkoxe4tky) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Tabular LLMs: An Introduction to the Foundation Models That Predict Your Spreadsheet","url":"https://daily.dev/posts/tabular-llms-an-introduction-to-the-foundation-models-that-predict-your-spreadsheet-dkoxe4tky","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/tabular-llms-an-introduction-to-the-foundation-models-that-predict-your-spreadsheet-dkoxe4tky"},"datePublished":"2026-07-24T16:41:32.030Z","dateModified":"2026-09-14T07:28:21.984Z","description":"Tabular foundation models — pretrained transformers that predict on any spreadsheet zero-shot via in-context learning — now outperform fully tuned...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/cdd8c41639e6fd76ea774e762abcc12e?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/cdd8c41639e6fd76ea774e762abcc12e?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Towards Data Science","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Towards Data Science","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/tds","url":"https://daily.dev/sources/tds"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/tabular-llms-an-introduction-to-the-foundation-models-that-predict-your-spreadsheet-dkoxe4tky","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"machine-learning","timeRequired":"PT23M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Towards Data Science","item":"https://daily.dev/sources/tds"},{"@type":"ListItem","position":3,"name":"Tabular LLMs: An Introduction to the Foundation Models That Predict Your Spreadsheet"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/tabular-llms-an-introduction-to-the-foundation-models-that-predict-your-spreadsheet-dkoxe4tky#faq","mainEntity":[{"@type":"Question","name":"Does TabICLv2 outperform tuned gradient-boosted trees like XGBoost or LightGBM on tabular data?","acceptedAnswer":{"@type":"Answer","text":"Yes, on the TabArena benchmark TabICLv2 sits 158 Elo above tuned-and-ensembled LightGBM (1432) despite requiring no hyperparameter tuning, and it wins 40 of 51 datasets against the best tuned-and-ensembled GBDT. It loses mainly on tables with more than 100 features or high-cardinality categorical columns, where trees remain the better choice. Anyone choosing between GBDTs and foundation models for tabular data can track these tradeoffs on daily.dev."}},{"@type":"Question","name":"How many parameters does TabICLv2 have and what license is it released under?","acceptedAnswer":{"@type":"Answer","text":"TabICLv2 has 27,552,258 parameters for its classifier checkpoint and 28,544,991 for its regressor, roughly a quarter the size of the t0-alpha time-series model. It is released as open weights on HuggingFace under a BSD-3 license with no gating or click-through requirement, and installs via pip install tabicl. Developers evaluating open-weight tabular models can follow licensing and size details like this on daily.dev."}},{"@type":"Question","name":"Are tabular foundation models faster or cheaper to run than tuned gradient-boosted trees?","acceptedAnswer":{"@type":"Answer","text":"TabICLv2's median fit time is about 4 seconds per 1,000 rows, roughly a hundred times shorter than tuned GBDT protocols, since fitting is a single forward pass with no hyperparameter search. Its median predict time, 0.38 seconds per 1,000 rows, beats tuned-and-ensembled LightGBM's 2.64 seconds, though it runs on GPU while default CatBoost remains the cheapest option on CPU at 0.08 s/1K. Teams weighing serving costs between foundation models and trees can compare benchmarks like this on daily.dev."}}]} ```