Skip to main content
CARLA HQ
PAIRWISE BENCHMARKAdaptive Repeated 3-Fold21 DatasetsMarkdown ↗

TabFM vs Kumo Tabular Small Benchmark

Empirical head-to-head evaluation of Google TabFM (Google Research) versus Kumo Tabular Small (Kumo AI / NVIDIA) on identical, precomputed IID or non-IID splits across 21 diverse tabular datasets.

HEAD-TO-HEAD WINS?Direct pairwise dataset matchup wins where the model achieved lower out-of-sample error loss (1 - AUC for binary, Log-Loss for multiclass, RMSE for regression).
TabFM11
:
6Kumo Tabular Small
(4 ties) across 21 benchmarks
TOURNAMENT ELO?Global Bradley-Terry Elo rating computed across all evaluation-split matchups with 1,000 bootstrap resamples. Higher is better (baseline 1000).
TabFM1106 (±99)
vs
Kumo Tabular Small1042 (±52)
Higher is better • 95% Bootstrap CI
MACRO IMPROVABILITY?Normalized average error regret relative to empirical suite ceilings across all 21 datasets. Lower is better (0.0% is optimal ceiling).
TabFM0.88%
vs
Kumo Tabular Small1.23%
Lower is better • Error regret vs ceiling
MEAN SPLIT LATENCY?Average execution latency in seconds per persisted IID or non-IID evaluation split across all benchmarks. Lower is faster.
TabFM742.64s
vs
Kumo Tabular Small14.52s
Lower is faster • adaptive split protocol
ARCHITECTURAL BLUEPRINTS

Model Specifications & Design Trade-offs

Comparing transformer architecture, Bayesian priors, sequence context windows, and browser execution runtime.

Google Research

Google TabFM

GitHub / Paper ↗

Google Research's tabular foundation model architecture focusing on robust columnar tokenization and zero-shot in-context transfer across diverse tabular benchmarks.

ArchitectureColumnar Embedding & Self-Supervised In-Context Transformer
Prior / PretrainingGenerative synthetic distributions and masked column modeling
Context WindowColumnar token budget per batch
Runtime PlatformServer-side only (Requires Python / JAX / PyTorch environment)
LicenseOpen Source (Apache 2.0)
Kumo AI / NVIDIA

Kumo Tabular Small

GitHub / Paper ↗

Kumo AI and NVIDIA's in-context tabular foundation model, pre-trained on diverse structured datasets for zero-shot classification and quantile regression.

ArchitectureIn-Context Tabular Foundation Model with Cross-Attention
Prior / PretrainingRelational & Structured Data Priors (SDM Recipe)
Context WindowConfigurable context rows with tiled row & ICL inference
Runtime PlatformNative WebGPU & WASM Browser Runtime (via ONNX Runtime)
LicenseOpen Source (Apache 2.0)
DEEP DIVE REPORTON-DEVICE RUNTIME ANALYSIS

TabICLv2 vs TabPFN-3 vs TabFM vs EXAONE: Which Model Can Run in the Browser?

Evaluating model weight size (MB vs GB), VRAM requirements, Python server infrastructure overhead, licensing terms, and WebGPU client-side execution compatibility.

HEAD-TO-HEAD BREAKDOWN

21-Dataset Metric Evaluation Matrix

Empirical performance on deterministic, application-appropriate IID or non-IID splits: 1 - AUC for binary classification, Log-Loss for multiclass, and RMSE for regression (all metrics: lower is better).

DatasetDomainTaskRows
Splits?Repeats × folds. IID datasets use three folds with size-aware repeats, and the same row indices are used for every model.
Metric?Primary evaluation error loss: 1 - AUC for binary classification, Log-Loss for multiclass, RMSE for regression. All metrics: lower is better.
TabFM
Loss?Google TabFM primary out-of-sample error loss (lower is better).
Kumo Tabular Small
Loss?Kumo Tabular Small primary out-of-sample error loss (lower is better).
Margin (Δ)?Error loss difference. Negative (green) indicates TabFM lead; positive (red) indicates Kumo Tabular Small lead.
Winner?Model achieving lower out-of-sample error loss on this benchmark split.
TabFM
Latency?Mean inference runtime in seconds per evaluation split for TabFM.
Kumo Tabular Small
Latency?Mean inference runtime in seconds per evaluation split for Kumo Tabular Small.
Abalone ↗Agriculture, Forestry & FishingREG4,1773×3RMSE2.022.02+0.00Kumo Tabular Small147.89s1.75s
Adult ↗Economics & Public PolicyBIN48,8423×31 - AUC0.06770.0677±0.00Tie4421.56s96.87s
Airfoil Self Noise ↗UCIREG1,50310×3RMSE1.111.02+0.09Kumo Tabular Small43.06s0.73s
Amazon Employee Access ↗Information Technology & Enterprise SecurityBIN32,7693×31 - AUC0.12730.1392-0.0119TabFM3389.93s48.40s
Bank Marketing ↗Finance & BankingBIN45,2113×31 - AUC0.18400.1880-0.0040TabFM3095.77s85.00s
Blood Transfusion Service Center ↗Healthcare & BiomedicineBIN74820×31 - AUC0.24450.2452-0.0007TabFM10.38s0.50s
Breast W ↗Healthcare & Life SciencesBIN69920×31 - AUC0.00550.0055±0.00Tie10.88s0.53s
Car ↗Automotive & Fleet ManagementMULTI1,72810×3Log-Loss0.01080.0253-0.0145TabFM25.35s0.65s
Compas Two Years ↗Legal & Public SafetyBIN5,2783×31 - AUC0.26190.2617±0.00Tie107.91s2.74s
Credit G ↗Finance & BankingBIN1,00010×31 - AUC0.19380.1956-0.0018TabFM19.29s0.77s
Diabetes ↗Healthcare & BiomedicineBIN76810×31 - AUC0.16230.1621+0.0003Kumo Tabular Small11.50s0.54s
Employee Salaries ↗Human Resources & Workforce AnalyticsREG9,2283×3RMSE10,4359,029+1,406Kumo Tabular Small466.77s5.53s
Fitness Club ↗Fitness, Sports & RecreationBIN1,50010×31 - AUC0.17950.1792+0.0003Kumo Tabular Small22.13s0.76s
House Sales ↗Real Estate & Property ValuationREG21,6133×3RMSE98,88798,512+374Kumo Tabular Small1839.00s26.72s
Houses ↗Real Estate & Urban PlanningREG20,6403×3RMSE37,48137,670-189TabFM1513.45s19.70s
Monks Problems 2 ↗Data Science & Artificial IntelligenceBIN60120×31 - AUC0.00000.0000±0.00Tie9.21s0.50s
Phoneme ↗Speech Processing & Acoustic EngineeringBIN5,4043×31 - AUC0.02020.0246-0.0043TabFM97.68s2.29s
Spambase ↗Cybersecurity & IT InfrastructureBIN4,6013×31 - AUC0.00660.0073-0.0008TabFM154.06s4.98s
Telco Customer Churn ↗Telecommunications & Subscription ServicesBIN7,0433×31 - AUC0.14740.1595-0.0120TabFM171.54s4.67s
Titanic ↗Maritime Safety & Actuarial Risk AnalysisBIN1,30910×31 - AUC0.11660.1276-0.0110TabFM22.09s0.84s
Vehicle ↗Automotive Engineering & Computer VisionMULTI84610×3Log-Loss0.19860.2244-0.0258TabFM15.97s0.53s
ZERO-SERVER CLIENT-SIDE EXECUTION

Evaluate Foundation Models In Your Spreadsheets

Carla HQ brings TabICLv2 foundation models directly into Google Sheets via WebGPU. Run zero-trust predictions on private data without API keys or cloud pipelines.

FREQUENTLY ASKED QUESTIONS

Frequently Asked Questions: TabFM vs Kumo Tabular Small

Common architectural and benchmarking questions comparing Google TabFM and Kumo Tabular Small.

How does TabFM compare to Kumo Tabular Small on tabular benchmarks?

Across the Carla benchmark suite, Google TabFM and Kumo Tabular Small are evaluated head-to-head on identical, precomputed IID or non-IID splits using standardized error losses (1 - AUC for binary classification, Log-Loss for multiclass, and RMSE for regression). IID datasets use adaptive repeated 3-fold evaluation to reduce variance. TabICLv2 enables 100% private in-browser WebGPU execution, whereas server-side models like Kumo Tabular Small require heavy PyTorch infrastructure and cloud GPU backend servers.

Can TabFM or Kumo Tabular Small run in the browser without server infrastructure?

Only TabICLv2 (via Carla HQ and nanotabicl ONNX Runtime WASM/WebGPU) is engineered to execute entirely client-side inside web browsers like Google Chrome. Google TabFM requires full Python and PyTorch server infrastructure with high memory allocations.

What are the primary architectural differences between TabFM and Kumo Tabular Small?

Google TabFM uses columnar embedding & self-supervised in-context transformer, whereas Kumo Tabular Small leverages in-context tabular foundation model with cross-attention. TabICLv2 specifically utilizes prefix attention with KV cache reuse for real-time tabular evaluation in spreadsheets.