---
modelA: "TabICLv2"
modelB: "Google TabFM"
canonicalSlug: "tabiclv2-vs-tabfm"
winsA: 2
winsB: 16
ties: 3
winRateA: "9.5%"
winRateB: "76.2%"
eloA: "932 (±62)"
eloB: "1111 (±87)"
avgLatencyA: "4.26"
avgLatencyB: "742.64"
canonicalUrl: "https://carlahq.eu/datasets/comparison/tabiclv2-vs-tabfm"
---

# TabICLv2 vs TabFM Tabular Benchmark Comparison

> Empirical head-to-head evaluation across Carla's open-source tabular foundation model benchmark suite.

## Head-to-Head Scoreboard
| Metric | TabICLv2 | Google TabFM | Comparison Result |
| :--- | :--- | :--- | :--- |
| **Head-to-Head Wins** | **2** (9.5%) | 16 (76.2%) | TabFM leads (3 ties) |
| **Bradley-Terry Elo Rating** | **932 (±62)** | 1111 (±87) | Global tournament rating across repeated splits |
| **Improvability Regret** | 2.11% | 0.85% | Normalized average error against dataset optimal ceilings |
| **Average Inference Latency** | 4.26 | 742.64 | Measured per-split forward-pass timing |
| **Browser Execution** | Native WebGPU Browser Runtime (Carla Engine / ONNX Runtime WASM) | Server-side only (Requires Python / JAX / PyTorch environment) | — |
| **Open Source License** | Open Source (BSD-3-Clause) | Open Source (Apache 2.0) | — |

## Architectural Specifications
| Dimension | TabICLv2 | Google TabFM |
| :--- | :--- | :--- |
| **Organization** | Inria SODA | Google Research |
| **Architecture** | In-Context Tabular Transformer with Prefix Attention & KV Caching | Columnar Embedding & Self-Supervised In-Context Transformer |
| **Pretraining Prior** | Synthetic Copula-based Quantile Priors & Prior.py DAGs | Generative synthetic distributions and masked column modeling |
| **Context Window** | 1,000+ context rows with zero fine-tuning required | Columnar token budget per batch |
| **Execution Engine** | Native WebGPU Browser Runtime (Carla Engine / ONNX Runtime WASM) | Server-side only (Requires Python / JAX / PyTorch environment) |
| **Repository** | [TabICLv2 Code](https://github.com/soda-inria/tabicl) | [TabFM Code](https://github.com/google-research/tabfm) |
| **Paper / ArXiv** | [TabICLv2 Paper](https://arxiv.org/abs/2602.11139) | [TabFM Paper](https://github.com/google-research/tabfm) |

## Dataset-by-Dataset Benchmark Results
| Dataset | Business Domain | Task Type | Metric | TabICLv2 Score | TabFM Score | Delta (%) | Winner |
| :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- |
| [Abalone](/datasets/undefined.md) | Agriculture, Forestry & Fishing | `regression` | undefined | 2.03 | 2.02 | +0.0% | **TabFM** |
| [Adult](/datasets/undefined.md) | Economics & Public Policy | `binary_classification` | undefined | 0.0793 | 0.0677 | +0.0% | **TabFM** |
| [Airfoil Self Noise](/datasets/undefined.md) | UCI | `regression` | undefined | 1.12 | 1.11 | +0.0% | **TabFM** |
| [Amazon Employee Access](/datasets/undefined.md) | Information Technology & Enterprise Security | `binary_classification` | undefined | 0.1483 | 0.1273 | +0.0% | **TabFM** |
| [Bank Marketing](/datasets/undefined.md) | Finance & Banking | `binary_classification` | undefined | 0.1970 | 0.1840 | +0.0% | **TabFM** |
| [Blood Transfusion Service Center](/datasets/undefined.md) | Healthcare & Biomedicine | `binary_classification` | undefined | 0.2446 | 0.2445 | +0.0% | **Tie** |
| [Breast W](/datasets/undefined.md) | Healthcare & Life Sciences | `binary_classification` | undefined | 0.0052 | 0.0055 | -0.0% | **TabICLv2** |
| [Car](/datasets/undefined.md) | Automotive & Fleet Management | `multiclass_classification` | undefined | 0.0231 | 0.0108 | +0.0% | **TabFM** |
| [Compas Two Years](/datasets/undefined.md) | Legal & Public Safety | `binary_classification` | undefined | 0.2701 | 0.2619 | +0.0% | **TabFM** |
| [Credit G](/datasets/undefined.md) | Finance & Banking | `binary_classification` | undefined | 0.2016 | 0.1938 | +0.0% | **TabFM** |
| [Diabetes](/datasets/undefined.md) | Healthcare & Biomedicine | `binary_classification` | undefined | 0.1632 | 0.1623 | +0.0% | **TabFM** |
| [Employee Salaries](/datasets/undefined.md) | Human Resources & Workforce Analytics | `regression` | undefined | 8,126 | 10,435 | -2308.8% | **TabICLv2** |
| [Fitness Club](/datasets/undefined.md) | Fitness, Sports & Recreation | `binary_classification` | undefined | 0.1795 | 0.1795 | -0.0% | **Tie** |
| [House Sales](/datasets/undefined.md) | Real Estate & Property Valuation | `regression` | undefined | 111,524 | 98,887 | +12637.0% | **TabFM** |
| [Houses](/datasets/undefined.md) | Real Estate & Urban Planning | `regression` | undefined | 40,971 | 37,481 | +3489.9% | **TabFM** |
| [Monks Problems 2](/datasets/undefined.md) | Data Science & Artificial Intelligence | `binary_classification` | undefined | 0.0000 | 0.0000 | 0.0% | **Tie** |
| [Phoneme](/datasets/undefined.md) | Speech Processing & Acoustic Engineering | `binary_classification` | undefined | 0.0278 | 0.0202 | +0.0% | **TabFM** |
| [Spambase](/datasets/undefined.md) | Cybersecurity & IT Infrastructure | `binary_classification` | undefined | 0.0075 | 0.0066 | +0.0% | **TabFM** |
| [Telco Customer Churn](/datasets/undefined.md) | Telecommunications & Subscription Services | `binary_classification` | undefined | 0.1488 | 0.1474 | +0.0% | **TabFM** |
| [Titanic](/datasets/undefined.md) | Maritime Safety & Actuarial Risk Analysis | `binary_classification` | undefined | 0.1229 | 0.1166 | +0.0% | **TabFM** |
| [Vehicle](/datasets/undefined.md) | Automotive Engineering & Computer Vision | `multiclass_classification` | undefined | 0.2401 | 0.1986 | +0.0% | **TabFM** |

## Canonical Resources
- **HTML Benchmark Page:** [https://carlahq.eu/datasets/comparison/tabiclv2-vs-tabfm](https://carlahq.eu/datasets/comparison/tabiclv2-vs-tabfm)
- **Carla Tabular Suite:** [https://carlahq.eu/datasets/](https://carlahq.eu/datasets/)