Summary: On November 10, 2026, Amazon introduced Mitra-v2, a tabular foundation model pretrained entirely on synthetic prediction problems — no real or customer data. By enlarging the pretraining tables tenfold in rows and roughly threefold in columns, and adding a richer synthetic prior, it now leads the single-model field on the TabArena benchmark. The classifier, regressor, and fine-tuning code are open under Apache-2.0 on Hugging Face, which matters if you’re tired of re-selecting algorithms and re-tuning hyperparameters for every new business table.
Why tabular prediction needed a different approach
Most of the data businesses run on lives in tables — rows of customers, transactions, or interactions, with attributes in columns. The standard tooling, gradient-boosted decision trees, builds a model tailored to one table at a time. Each new table means picking an algorithm, engineering features, tuning hyperparameters, and wiring up a pipeline again.
Tabular foundation models attack that cycle the way LLMs attack text: pretrain once on millions of synthetic prediction problems, then at deployment treat the labeled rows of a previously unseen table as the examples in context and predict the unlabeled rows. The original Mitra worked well on small and mid-sized tables but ran out of room on large tables, wide tables, and problems with many classes. Mitra-v2 is the fix, led by Amazon Connect Customer, which is itself adopting the model for prediction on structured business and customer-experience data.
This category is getting real money behind it — as I wrote when SAP spent €1B on Prior Labs, tabular foundation models are now strategic. Mitra-v2 is the open-weights counterweight worth testing against anything you’re paying for.
What actually changed under the hood
The architecture is untouched: a Transformer that attends across the rows of a table and across its columns, carried over from Mitra-v1. The gains come from two pretraining decisions.
First, scale of the examples. Mitra-v2 sees pretraining tables with 10 times as many rows and roughly 3 times as many columns as Mitra-v1 did, so the model has now encountered tables shaped like the ones businesses actually use.
Second, a new family of synthetic tables. The original Mitra generated data from structural causal models — graphs of cause-and-effect relationships between columns — and from tree-based generators. Mitra-v2’s hybrid SCM prior builds a single causal graph in which each relationship can be a small neural network, a decision tree, a smooth function drawn from a Gaussian process, or a time-dependent autoregressive process. One synthetic table can therefore carry the mix of relationship types you find in real data, and a new hybrid optimizer keeps training stable across that wider range.
Every pretraining table is synthetic. The authors emphasize that no real data — and in particular no customer data — is used; the model learns how columns relate to each other in general, not the contents of any real dataset.
Wrappers and the bagging caveat
Two lightweight wrappers extend the model past its pretraining limits. For very wide tables, singular-value decomposition compresses the data into a smaller space that captures most of the variance. The output layer has only 10 nodes, matching 10 classes, so for tasks with more classes than that, each node’s output feeds a hierarchical tree that refines the classification further.
Here’s the detail worth pausing on: by default, Mitra-v2 is briefly and automatically fine-tuned on each new table it sees, producing eight copies whose predictions are averaged — bagging, a standard ensemble technique. So despite the zero-shot framing of tabular foundation models generally, this isn’t quite a zero-shot model. The adaptation adds computational overhead in exchange for higher accuracy. The blog doesn’t quantify that overhead, so if inference cost is tight for you, that’s the first number to measure yourself.
The benchmark picture
On TabArena, a public benchmark of 51 real-world datasets maintained by an independent academic team and ranked by Elo ratings computed from head-to-head wins under a one-hour budget, Mitra-v2 leads the single-model field. It beats Mitra-v1 on 32 of the 33 datasets where both models run natively, and it completes all 51 datasets within the one-hour budget — Mitra-v1 failed to do that on more than a third of them. It also ranks among the leading methods on the independent TALENT benchmark of roughly 300 datasets, strongest on classification tasks with more than 10 classes.
For regression, Mitra-v2 predicts a range of likely values per row rather than a single number, so you get prediction intervals along with the point estimate. On ScoringBench, which evaluates exactly that — forecasts that are accurate and well calibrated — the fine-tuned model ranks first.
What the blog does not establish: latency and cost figures in production, behavior on tables that exceed the wrappers’ compression capacity, or how its economics compare against simply running gradient boosting per table. The TALENT result is summarized without per-dataset specifics; the technical report is where the details live.
What to do with this
The takeaway is that synthetic pretraining quality and diversity — not a bigger network — drove the step change here. If you own tabular prediction workloads, the practical next step is a bake-off: pick two or three of your real tables, run the open Mitra-v2 classifier and regressor within a one-hour budget mirroring TabArena’s, and compare against your current gradient-boosted pipeline on both accuracy and wall-clock cost. Watch two failure modes specifically: tables too wide for the SVD wrapper to preserve signal, and class counts far above 10 where the hierarchical tree wrapper carries the load. Mitra-v2 gives you a free Apache-2.0 baseline before you commit to anyone’s commercial offering.
Sources
View the original post on X