TabICL ran up to 10 times faster in Databricks model comparison
Databricks' five-model repository shows TabICL up to 10 times faster than TabPFNv2, while flagging GPU costs, restrictive weights and the case for fine-tuning.
By Nathan Brooks
4 min read

What's News
- TabICL ran up to 10 times faster than TabPFNv2 while matching its classification accuracy.
- The TabICL evaluation covered up to 500,000 samples, 500 features and 53 datasets with more than 10,000 samples.
- Databricks published the guide on November 16, 2026 and introduced AI Runtime in public preview on March 19, 2026.
- On-demand H100 GPUs can support TabPFN-3 at up to 1 million rows and 200 features.
- TabFM and TabPFNv3.5 weights are non-commercial, while TabICLv2, Mitra v2 and KumoTabular have permissive licenses.
By Nathan Brooks
TabICL ran as much as 10 times faster than TabPFNv2 while matching classification-benchmark accuracy in a model comparison published by Databricks.
Databricks published its technical guide on November 16, 2026, alongside a public repository for comparing five model providers and testing fine-tuning workflows.
The public solution accelerator includes notebooks and a shared test harness for TabPFN, TabFM, TabICL, Nori and KumoTabular.
Databricks targets persistent table-prediction workloads:
- Credit-risk predictions
- Customer-churn predictions
- Demand forecasts
- Predictive-maintenance models
Under conventional development, each task can require a separate training and retraining pipeline, according to the guide.
What does this change for enterprise AI teams?
The change is operational, not merely a benchmark result. Databricks positions the repository as a way to test whether tabular foundation models can reduce per-task machine-learning work without moving data into separate GPU tools.
According to the guide, tabular foundation models arrive pretrained on millions of diverse synthetic datasets. In zero-shot inference, labeled rows become context for rows needing predictions. Transformer attention relates new rows to those examples and returns predictions in one forward pass without parameter updates.
Teams can still fine-tune the pretrained weights on proprietary data when zero-shot performance falls short, Databricks says.
TabPFNv2 alternates attention within rows and columns to learn interactions, according to Databricks. Its computational cost rises rapidly as the number of samples and features increases.
The TabICL research paper takes a different route, building fixed-dimensional row embeddings through column-then-row attention before dataset-level in-context learning. The evaluated version handles up to 500,000 samples and 500 features.
The paper’s authors report matching TabPFNv2 on their classification benchmark while running as much as 10 times faster. TabICL also surpassed both TabPFNv2 and CatBoost on 53 datasets containing more than 10,000 samples.
That evaluation covered classification only. The holdout-tuned, single-model results also failed to capture potential gains from cross-validated ensembles, the researchers said.
Licensing creates a separate deployment filter, and Databricks’ guide divides the models by their commercial terms:
- TabICLv2 uses BSD 3-Clause, which the guide classifies as commercially permissive.
- Mitra v2 uses Apache-2.0, also commercially permissive.
- KumoTabular uses OpenMDW-1.1, likewise commercially permissive.
- TabFM and TabPFNv3.5 weights are restricted to non-commercial use; commercial deployment requires agreements with their creators.
- The repository distinguishes TabPFN’s commercial hosted API from its non-commercial weights.
Memory remains an operational constraint even with row compression. Because in-context learning supplies the training set with every prediction, GPU memory demand grows with the number of rows and features. Databricks recommends testing accuracy, inference latency, memory use and cost on an organization’s own data.
Databricks says GPU inference costs more per prediction than running XGBoost on a CPU. The potential saving is engineering and maintenance time, while the continuing expense is accelerated compute.
The reference implementation gives model comparisons a common basis. Every supported implementation reads the same synthetic retail and consumer-packaged-goods datasets and uses identical training and test splits. The repository logs common metrics and scikit-learn baselines in one results table.
Examples cover classification, regression, outlier detection and lag-based forecasting where each model supports them, Databricks says.
Compute requirements differ by deployment. Self-hosted TabFM, TabICL, Nori and KumoTabular can run on serverless CPUs, although the repository recommends GPUs for practical performance. Its TabPFN example calls a hosted API that works with CPUs, while a TabICL example lets CPU pipelines offload inference to a dedicated GPU serving endpoint.
The production path
Databricks introduced AI Runtime in public preview on March 19, 2026, as a serverless stack for distributed GPU training. On-demand H100 GPUs can support TabPFN-3 at up to 1 million rows by 200 features, according to Databricks.
AI Runtime bundles notebooks, Lakeflow Jobs, MLflow experiment tracking and GPU-utilization observability. It also provides native Docker support and remote SSH through the Databricks CLI.
A KumoTabular fine-tuning demonstration exposes GPU usage and MLflow metrics inside a notebook, then registers the resulting model in Unity Catalog. For production, Databricks describes high-throughput batch pipelines or autoscaling real-time REST endpoints through Model Serving. Unity Catalog manages model versions and governance.
Databricks says the business case ultimately depends on accuracy across an organization’s own tables, usable license terms and whether lower engineering overhead offsets GPU-backed prediction costs. The next purchasing decision is therefore measurable: which model clears corporate accuracy, latency, memory and licensing thresholds without letting accelerated compute erase the engineering savings.
Source: unite.ai
More from Nathan Brooks
Show full bio
News editor covering regulation, policy and their impact on entrepreneurship, startups and venture funding at Business Bearings.
2,232 articles