Indian AI model aims to forecast numbers that change over time

Birla AI Labs, an AI initiative and venture studio under Aditya Birla Group, has released its first time series foundation model. It is called Yug (Analytics India Magazine calls it Yug-1).

A time series is a set of numbers measured over and over across time, for example daily sales in a store, electricity use or traffic on a road. A foundation model is a large AI model trained broadly so it can be used for many tasks. Yug is meant to forecast how such numbers will develop.

How it works

According to the model card on Hugging Face, the model has 271.8 million parameters, which are roughly the settings the model learns during training. It produces probabilistic forecasts without task-specific training (zero-shot). That means it can be used on new data without first being trained specifically for it. Instead of a single estimate, it gives nine levels from 0.1 to 0.9, so users can see a range of possible outcomes. It can take at most 2,048 previous data points into account as background.

The model can make forecasts both from a single data series and from series with additional influencing factors, known as covariates. However, the model card notes that forecasts with covariates degrade faster the further ahead you look. For long horizons, the simpler approach without covariates is often the better choice.

The model was trained on parts of the public datasets Chronos Datasets and GIFT-Eval Pretrain, as well as on the company's own synthetic data.

Performance is the company's own claim

The model card lists a CRPS of 0.529 and a MASE of 0.744 on the GIFT-Eval benchmark. These are the developers' own figures. Ananya Birla, founder and chair of Birla AI Labs, says Yug-1 performed better than other first-generation models and competitively against newer ones. The sources do not indicate that anyone else has verified this.

The company calls the model "Asia's first time series foundation model built for real-world use." That is the company's own description.

Use and license

The use cases mentioned are demand forecasting, financial analysis, equipment monitoring and infrastructure planning. Ananya Birla says: "Most of the conversation around AI today is about language. But so much of the data that actually runs businesses and economies moves through time."

The model can be downloaded from Hugging Face and GitHub and installed with "pip install yug." The model weights are licensed under CC-BY-NC 4.0, which permits only non-commercial use. According to PyPI, the code is under Apache-2.0. For businesses, this means the model itself cannot simply be used commercially.

According to Analytics India Magazine, Birla AI Labs has a presentation planned at NeurIPS 2026.