Amazon Open-Sources Mitra-v2, a Foundation Model That Predicts Missing Values in Business Tables Without Building a New Model Each Time

Most of the data that runs a business does not live in chat logs or images. It lives in tables: customers in rows, their attributes in columns, and a long list of questions that teams want answered, such as who is likely to churn, which order will be returned, or how much demand to expect next week. On November 10, 2026, Amazon released Mitra-v2, the second generation of its tabular foundation model, and made it free to use under the Apache-2.0 license. The pitch is simple and potentially disruptive: one pretrained model that can read a table it has never seen and predict the missing values, without a team building a new model from scratch each time.

What Amazon Released

According to the Amazon Science announcement, Mitra-v2 comes in versions fine-tuned for both classification and regression, and Amazon has published the models together with their inference and fine-tuning code on Hugging Face. The work was led by Amazon Connect Customer, the unit that builds agentic-AI systems for core business functions, and that team says it is adopting Mitra-v2 to make predictions on structured business and customer-experience data, for example estimating how likely an important event is given a customer’s interaction history and operational context.

The first Mitra, introduced last year, handled small and mid-sized tables well but struggled with large tables, wide tables and problems with many classes. The second version targets exactly those gaps.

How a Tabular Foundation Model Works

For years, the default tool for table predictions has been gradient-boosted decision trees. They are accurate, but every new table means a fresh project: choosing an algorithm, engineering features, tuning hyperparameters and maintaining a pipeline. Tabular foundation models borrow the trick that made large language models useful. A language model can learn from a few examples in a prompt; Mitra-v2 treats the labeled rows of a real table as those examples and predicts the unlabeled rows, a technique known as in-context learning.

One detail matters a great deal for enterprise buyers. Amazon says every table used to pretrain Mitra-v2 is synthetic, and that no real data, and in particular no customer data, went into pretraining. The model is meant to learn general rules about how columns relate to one another rather than memorize any real dataset.

Bigger Synthetic Tables, Same Architecture

Amazon did not make the network deeper. The architecture is still a Transformer that pays attention both across rows and across columns. What changed is the training diet. Mitra-v2 was pretrained on tables with 10 times as many rows and about three times as many columns as those used for Mitra-v1, bringing its training examples closer to the size of tables businesses actually use.

The team also added a new “hybrid” family of synthetic data. Instead of generating each table from a single kind of cause-and-effect graph or tree model, the new generator builds one causal graph in which each relationship can be a small neural network, a decision tree, a smooth Gaussian-process function or a time-dependent autoregressive process. The goal is a single synthetic table that mixes relationships the way messy real-world data does. A new hybrid optimizer keeps training stable across that wider range.

Two lightweight add-ons stretch the model at deployment time. For very wide tables, Amazon uses singular-value decomposition to compress the data into a table that fits while keeping most of its variance. Because the output layer has only 10 nodes, tasks with more than 10 classes are routed through a hierarchical tree that refines the answer. Amazon is also candid that Mitra-v2 is not purely zero-shot: by default it is briefly fine-tuned on each new table, producing eight copies whose predictions are averaged, which adds compute but improves accuracy.

The Benchmark Numbers

• TabArena: On this independent benchmark of 51 real-world datasets, Amazon says Mitra-v2 leads the single-model field. It beat Mitra-v1 on 32 of the 33 datasets where both run natively and finished all 51 within the benchmark’s one-hour budget, something Mitra-v1 failed to do on more than a third of them.
• TALENT: On this benchmark of roughly 300 datasets, Amazon reports Mitra-v2 ranks among the leading methods and is strongest on classification tasks with more than 10 classes.
• ScoringBench: For regression, Mitra-v2 predicts a range of likely values rather than one number, giving users prediction intervals. Amazon says the fine-tuned model ranks first on this leaderboard, which rewards forecasts that are both accurate and well calibrated.

These results come from Amazon’s own announcement and technical report, so they are worth validating on your own data before making decisions. Still, TabArena is maintained by an independent academic team, which makes the leaderboard claim easier to check than a purely internal test.

Why It Matters for Business

It lowers the cost of everyday prediction. Churn scoring, lead scoring, fraud flags, demand forecasts and support-ticket routing are some of the highest-return machine learning projects a company can run, yet each one usually needs a data scientist and weeks of feature work. A model that handles a new table with little more than its labeled rows could put these predictions within reach of smaller teams.

The license removes a common blocker. Apache-2.0 allows commercial use and modification, and the models can run on a company’s own infrastructure. Combined with pretraining on synthetic data only, that answers two questions legal and security teams ask first: can we use it, and whose data is inside it?

Uncertainty estimates make forecasts more useful. A demand forecast that says “probably 1,200 units, likely between 1,000 and 1,450” is far easier to plan inventory and staffing around than a single number. Regression with built-in prediction intervals is the kind of feature operations teams care about.

It shows where Amazon is heading. Because Amazon Connect Customer led the work and plans to use it on customer-experience data, Mitra-v2 hints that predictive scoring will increasingly be built into contact-center and customer-service platforms rather than bolted on by separate analytics teams.

There are trade-offs. Per-table fine-tuning and bagging add compute at prediction time, and gradient-boosted trees remain fast, well understood and hard to beat on many problems. The sensible move is a head-to-head test on one or two real business tables, comparing accuracy, cost and how much engineering time each approach takes.

The Bottom Line

Foundation models have transformed text and images, and Mitra-v2 is a serious attempt to bring the same “pretrain once, use everywhere” idea to the spreadsheets and databases where business value actually sits. Its main lesson is that richer, more varied synthetic training data, not a bigger network, drove the gains. With open weights, a permissive license and leaderboard results to back it up, it is worth a trial for any team that spends too much time rebuilding the same prediction models.

Source: Amazon Science, “Mitra-v2: Scaling synthetic pretraining for tabular foundation models” (November 10, 2026).

Edited by Coron @QUE.COM
Website: https://QUE.COM Intelligence
Sponsored by: https://MAJ.COM AI Autonomous