What happened
Amazon Science announced Mitra-v2 on November 10, 2026. The model is intended for classification and regression on tables. Rather than train a dedicated model for every table, it uses labeled rows from a new table as examples and predicts the unlabeled rows through in-context learning. Amazon says it has released classification and regression versions, along with inference and fine-tuning code, under the Apache-2.0 license on Hugging Face. The company also describes a technical report. Amazon Connect Customer led the effort and is adopting the model for predictions on structured business and customer-experience data.
Amazon reports results on two public benchmarks. On TabArena, which contains 51 real-world datasets, Mitra-v2 leads the single-model field. It performs better than Mitra-v1 on 32 of the 33 datasets where both models run natively, and completes all 51 within the benchmark's one-hour budget; Mitra-v1 did not complete more than a third within that budget. On ScoringBench, a benchmark for regression that predicts a range of likely outcomes, fine-tuned Mitra-v2 ranks first.
The architecture remains a Transformer that attends across both rows and columns. The pretraining scale changed: Mitra-v2 saw tables with ten times as many rows and roughly three times as many columns as Mitra-v1. Amazon also added a hybrid synthetic-data prior. It combines causal relationships in a graph, with each relationship represented by a small neural network, decision tree, Gaussian-process function, or time-dependent autoregressive process. Amazon says the pretraining uses synthetic tables, not real or customer data.
The model uses additional steps at deployment. Singular-value decomposition converts very wide tables into a form that fits its allotted input space. For classification with more than 10 classes, a hierarchical tree refines the output of its 10-node output layer. By default, the system briefly fine-tunes on each new table and averages predictions from eight copies using bagging. Amazon explicitly says this means the model is not quite zero-shot and that adaptation adds computational overhead.
What the archive adds
New Runtime's earlier coverage of synthetic-data generation as a designed system (October 21, 2026) focused on making data-generation workflows reproducible and auditable. Mitra-v2 is a new event, but it extends the same broader idea into tabular pretraining: the composition of synthetic examples is part of the model design, not merely a way to increase dataset volume.
Trend signal
This supports the direction in agent economics moves to completed work only weakly: Amazon's benchmark comparisons measure predictive performance and completion within a time budget, not the cost of verified completed work. The release does, however, make evaluation on task-level datasets more actionable than model-scale claims alone. It does not establish a production cost or quality advantage for builders.
Why now
Amazon's announcement frames Mitra-v2 as a response to limitations in the original Mitra on large and wide tables and classification tasks with many classes. Its larger synthetic training tables and added hybrid prior are aimed at adapting to more realistic business table sizes without training a separate model for each table.
Why it matters
Tabular prediction remains a common business task, but the conventional workflow requires teams to select an algorithm, engineer features, tune hyperparameters, and maintain a pipeline for each dataset. Mitra-v2 offers a different experiment: put labeled rows in context, then predict the remaining rows with released code and weights. That could reduce repeated model-building work if it proves competitive on a team's own data.
The benchmark results are a reason to test, not a reason to replace a production baseline. The default adaptation step incurs additional compute, and practical fit will depend on latency, table shape, class count, and data quality. Builders should compare it with their existing gradient-boosted trees or other methods using the same splits and operational constraints. The value is in making that comparison accessible under an open license, while keeping the model's non-zero-shot behavior explicit.
Limits and uncertainty
The reported benchmark rankings come from Amazon's announcement; this note does not independently reproduce them. The announcement describes the model's default adaptation and input-handling methods, but benchmark leadership alone does not establish performance across a particular company's data, deployment costs, latency, or production reliability. Amazon Connect Customer's adoption is an announced use, not a disclosed quantified customer outcome. Teams should inspect the technical report and released code before relying on the reported results.
Skill to build now
Take one tabular classification or regression problem with a stable held-out split. Run Mitra-v2 using the released inference code, then compare it with your current baseline on predictive quality, latency, and compute cost. Record whether automatic fine-tuning and any wide-table or multi-class handling are invoked; those steps change the comparison from a pure zero-shot test.
Reader takeaway
Mitra-v2 is an open, benchmark-leading tabular foundation model according to Amazon, but it is not a zero-shot shortcut: its default workflow adapts to each table and averages eight copies. Treat the release as a testable alternative to table-specific pipelines, and validate it against your own baseline before making it part of production.