Databricks Adds Tabular Foundation Models to AI Runtime
Start here
Databricks has added support for tabular foundation models (TFMs) on its AI Runtime, letting users experiment with and fine-tune pretrained transformer models for table prediction on managed GPU infrastructure. TFMs predict on new tables through in-context learning without task-specific training, according to Databricks. The company compared three models: TabICLv2 (about 28M parameters, BSD 3-Clause license), TabFM (about 400M parameters, non-commercial license) and TabPFNv3.5 (about 220M parameters). The licensing differences mean TabFM's non-commercial terms limit where it can be used, while TabICLv2's permissive license does not. Databricks frames the release as guidance on how TFMs work and when to adopt them, rather than a claim of measured performance gains.
Generated from reports · updated 1 h ago
Report timeline
Follow the reports to see the event from different sides.
- Databricks BlogTabular Foundation Models on Databricks AI Runtime: How TFMs Work and When to Adopt Them
Databricks examines tabular foundation models (TFMs), pretrained transformers that predict on new tables via in-context learning without task-specific training. It compares TabICLv2 (~28M parameters, BSD 3-Clause), TabFM (~400M, non-commercial license) and TabPFNv3.5 (~220M), and shows how Databricks AI Runtime supports TFM experimentation and fine-tuning with managed GPU infrastructure.
Heat for this event
Not enough continuous observations yet to draw a trend.