Amazon Science announced Mitra-v2 on November 10, 2026, as the second generation of its tabular foundation model, for classification and regression. The first Mitra was pretrained only on synthetic tables and handled small and mid-sized tables well, but struggled with large tables, wide tables and problems with many classes. Amazon Connect Customer led the work and is adopting the model for prediction on structured business and customer-experience data.
The release includes fine-tuned classification and regression versions, plus inference and fine-tuning code, all published on Hugging Face under Apache-2.0. Mitra-v2 leads the single-model group on TabArena, a public benchmark covering 51 real-world datasets. It beats Mitra-v1 on 32 of the 33 datasets both models can run natively, and it finishes all 51 datasets within the one-hour budget, which Mitra-v1 failed on more than a third of them.
Mitra-v2 keeps its Transformer design, but its pretraining tables have 10 times the rows and roughly three times the columns of Mitra-v1's. A new hybrid SCM prior mixes neural networks, decision trees, Gaussian process functions and autoregressive processes in one causal graph. Wide tables use singular-value decomposition, and tasks with more classes use a hierarchical tree. By default each new table gets brief fine-tuning, with eight averaged copies.
Key facts
- Mitra-v2 is the second generation of the tabular foundation model, for classification and regression.
- Mitra-v2 is pretrained on millions of synthetic prediction problems, and no real data is used.
- Mitra-v1 failed the one-hour budget on more than a third of TabArena's 51 datasets.
- On ScoringBench, the fine-tuned Mitra-v2 model places first among regression entries.
- The output layer has 10 nodes, so tasks with more classes get a hierarchical tree.
Why it matters
The post is dated November 10, 2026. The fine-tuned classification and regression models and the code are open on Hugging Face under Apache-2.0, and the team invites researchers and practitioners to try them on their own tables. Amazon Connect Customer is taking up Mitra-v2 to make predictions from structured business and customer-experience data, such as judging the chance that a business-relevant event happens.




