Amazon releases Mitra-v2, an open tabular foundation model that leads the single-model field on TabArena

Amazon has released Mitra-v2, a second-generation tabular foundation model for classification and regression that leads the single-model field on TabArena, according to Amazon Science. Fine-tuned classification and regression versions, plus inference and fine-tuning code, are available on Hugging Face under the Apache-2.0 license. Amazon describes the work in a technical report and a blog post by Chris (Luyang) Kong.

What changed from Mitra-v1

The architecture is unchanged: a Transformer that attends both across the rows and across the columns of a table. The gains come from the synthetic data used in pretraining.

• Larger pretraining tables: Mitra-v2 is pretrained on tables with 10 times as many rows and roughly three times as many columns as those used for Mitra-v1.
• A new hybrid SCM prior: The original Mitra used tables generated from structural causal models (SCMs) and from tree-based models. The new prior builds a single causal graph in which each relationship can be a small neural network, a decision tree, a Gaussian-process function, or a time-dependent autoregressive process.
• A hybrid optimizer: Amazon says it keeps training stable across the wider range of synthetic data.

Every pretraining table is synthetic. Amazon states that no real data, and no customer data, is used.

Deployment-time wrappers

Two lightweight wrappers extend the model beyond its pretraining limits. For very wide tables, singular-value decomposition produces a smaller table that Amazon says captures most of the variance in the original. Mitra-v2's output layer has 10 nodes, one per class, so for tasks with more than 10 classes each node's output is sent to a hierarchical tree that refines the classification.

By default, Mitra-v2 is briefly and automatically fine-tuned on each new table, producing eight copies whose predictions are averaged (bagging). Amazon says this means it is not strictly a zero-shot model. The adaptation adds computational overhead in exchange for higher accuracy.

Benchmark results

All results below are reported by Amazon.

• TabArena: On this public benchmark of 51 real-world datasets, maintained by an independent academic team and ranked by Elo under a fixed one-hour budget, Mitra-v2 leads the single-model field. The Elo comparison is dated September 2026.
• Versus Mitra-v1: Mitra-v2 produces a better result on 32 of the 33 datasets where both models run natively. It completes all 51 datasets within the one-hour budget, which Mitra-v1 failed to do on more than a third of them.
• TALENT: On this independent benchmark of roughly 300 datasets, Mitra-v2 ranks among the leading methods and is strongest on classification tasks with more than 10 classes.
• ScoringBench: On this public regression benchmark, which rewards forecasts that are both accurate and well calibrated, fine-tuned Mitra-v2 ranks first.

For regression, Mitra-v2 predicts a full range of likely values for each row, giving prediction intervals along with the point estimate. Numeric Elo values are shown only in charts in the post, and the technical report contains per-dataset comparisons.

Adoption

Amazon Connect Customer, which provides agentic-AI systems for core business functions, led the effort. It is adopting Mitra-v2 for prediction on structured business and customer-experience data, such as estimating how likely a business-relevant event is to occur given a customer's interaction history and operational context.

What this means

Amazon attributes the improvement to the diversity and range of the synthetic pretraining data rather than to a deeper network, since the architecture is the same as Mitra-v1. For teams working with tabular data, the Apache-2.0 release of both the fine-tuned models and the fine-tuning code means Mitra-v2 can be tested directly against per-table pipelines such as gradient-boosted decision trees. The default per-table fine-tuning and eight-copy bagging add compute, so cost and latency are worth measuring against a team's own tables. The strongest reported results are on tables with more than 10 classes (TALENT) and on calibrated regression (ScoringBench), where Mitra-v1 had run out of room, according to Amazon.