Amazon releases Mitra-v2, an open model that predicts from tables

A team from Amazon Connect Customer, which makes agentic AI systems for firms, released Mitra-v2. It is a model that reads the labeled rows of a new table and predicts the missing values with no training run. The team released the code and models for free with the Apache-2.0 license on Hugging Face. Amazon Connect Customer will use it for customer data.

How it works

The team used only synthetic tables as the data for Mitra-v2, and it used no customer data. The tables had 10 times as many rows and about 3 times as many columns as for Mitra-v1. By default, the model also tunes for a short time on each new table.

The numbers

On TabArena, a public test of 51 datasets, Mitra-v2 had a score above Mitra-v1 on 32 of 33 datasets. These are the datasets where the 2 models work. It did all 51 datasets in the 1 hour limit. Mitra-v1 did not do this on more than a third of them. For regression, the tuned model is first on ScoringBench.

What is not known

The tuning adds computing cost, and the source gives no numbers for it. The source gives no results from the use at Amazon Connect Customer. The authors wrote the report, and an external team keeps TabArena.

Insight: Amazon gives away the model and keeps the product

Mitra-v2 uses only synthetic data, thus no customer data is in it. It has the Apache-2.0 license. In October, Mistral released Large 4 with a license. A firm must have a deal with Mistral to sell access to the model. This reads as a move: a free model with a clear license gets wide use, while Amazon keeps the tool that uses it. Builders can test it free on their tables.