Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
InfrastructureInsight · On RadarTNW→ original

Databricks shows how AI forecasts from tables without training: up to 1 million rows

Usually, each tabular task, such as credit risk or customer churn, gets its own trained model. Databricks has published code for five AI models that adapt to your examples at the moment of prediction. On Nvidia H100 GPUs, TabPFN-3 handles up to 1 million rows, but each prediction costs more than the XGBoost boosting model on a regular CPU. Before deployment, compare the models on your own data.

AI-processed from TNW; edited by Hamidun News
Databricks shows how AI forecasts from tables without training: up to 1 million rows
Source: TNW. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTNW ↗
Editor-in-Chief
Key metric #1TabICLv2, Mitra v2 and KumoTabular can be used commercially
Key metric #2TabFM and the new TabPFN require a separate deal with their authors
Memory grows with the number of rowsthe model rereads the entire sample

Imagine a bank needs to assess credit risk on loan applications, while a store wants to know which customers will leave soon. This is tabular data: rows are customers, and columns are features such as age, purchase totals or the number of missed payments. Typically, a team trains a separate model for each such task, most often the XGBoost boosting model, and then retrains it regularly.

On Monday, Databricks demonstrated a different approach: AI models that make predictions from a table without separate training for the task. The company released a guide and open code for five such models. According to Databricks, one of them, TabPFN-3, works with tables of up to 1 million rows and 200 columns on Nvidia H100 GPUs.

How it works in plain terms

These models are called tabular foundation models. They are pre-trained on millions of synthetic tables so they learn general patterns: how columns typically relate, and what noise and outliers look like. When a company brings its own table, nothing needs to be fine-tuned. You pass the model example rows with known outcomes, such as past customers who left or stayed, and it adapts to them at the moment of prediction.

This resembles an experienced analyst. They don't build a new model from scratch; they look at your past cases and immediately evaluate a new one. Databricks published code that compares five models: TabICLv2, Mitra v2, KumoTabular, TabFM and the new TabPFN. The models are meant to run on AI Runtime, Databricks' pay-as-you-go GPU service, which has been in preview since March.

The direction is gaining traction. SAP, which works closely with Databricks, is collaborating with Prior Labs, the creator of TabPFN, and H2O.ai has also released its own development in this area.

The catch

Databricks itself names the drawbacks, and they are significant. The first is cost. A single prediction on a GPU costs more than classic XGBoost on a regular CPU. If you run millions of predictions a day, the difference quickly becomes a noticeable line item on the bill.

The second is memory. The model rereads the entire training sample on every prediction, so memory consumption grows with the number of rows. With ordinary boosting, after training, the situation is different: the finished model is small and fast.

The third is licensing. TabICLv2, Mitra v2 and KumoTabular can be used commercially. TabFM and the newest version of TabPFN, however, must be obtained under a separate agreement with their authors. If you planned to adopt the latest and most powerful model right away, check the terms first.

Databricks itself advises not taking promises at face value and first testing the models on your own data. The publication does not imply that tabular foundation models outperform boosting in accuracy everywhere. This is a new tool with its own strengths, not a replacement for everything that worked before.

What practitioners should do

The main advantage of this approach is speed to start. If you need to test a hypothesis quickly, build a prototype, or tackle a task without the resources for a full ML cycle, a ready-made model that requires no training can save weeks. This is especially useful where labeled data is scarce and building a retraining pipeline for a single task makes no sense.

For high-volume production with a tight budget per prediction, XGBoost and its CPU-based analogs still look more reliable and cheaper. A reasonable scheme is to prototype on a tabular foundation model, compare its quality against boosting on the same data, and then decide whether paying for GPUs on an ongoing basis is worth it.

Databricks' open code lets you compare five models on one bench without assembling the infrastructure yourself. For teams already on Databricks, this is a quick way to see whether the idea pays off on their tables. Everyone else should review the licensing restrictions and the cost per prediction before building such a model into their architecture.

Story

Key Questions Answered

How it works in plain terms?

These models are called tabular foundation models. They are pre-trained on millions of synthetic tables so they learn general patterns: how columns typically relate, and what noise and outliers look like. When a company brings its own table, nothing needs to be fine-tuned. You pass the model example rows with known outcomes, such as past customers who left or stayed, and it adapts to them at the moment of prediction. This resembles an experienced analyst.

What are the key details on The catch?

Databricks itself names the drawbacks, and they are significant. The first is cost. A single prediction on a GPU costs more than classic XGBoost on a regular CPU. If you run millions of predictions a day, the difference quickly becomes a noticeable line item on the bill. The second is memory. The model rereads the entire training sample on every prediction, so memory consumption grows with the number of rows. With ordinary boosting, after training, the situation is different: the finished model is small and fast.

What practitioners should do?

The main advantage of this approach is speed to start. If you need to test a hypothesis quickly, build a prototype, or tackle a task without the resources for a full ML cycle, a ready-made model that requires no training can save weeks. This is especially useful where labeled data is scarce and building a retraining pipeline for a single task makes no sense. For high-volume production with a tight budget per prediction, XGBoost and its CPU-based analogs still look more reliable and cheaper.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…