Most of the conversation about artificial intelligence today is about language — chatbots that write, translate and answer questions. But the data that actually runs businesses and economies rarely arrives as sentences. It arrives as numbers over time: sales rising and falling with the seasons, machines humming a little differently before they fail, power demand climbing on a hot afternoon, prices shifting day after day. In India, a new model has been built to read exactly that kind of data.
That model is Yug-1, the first time-series foundation model from Birla AI Labs, the artificial intelligence initiative and venture studio under the Aditya Birla Group. Released in October 2026, Yug-1 is designed to forecast outcomes, find patterns and detect anomalies across interconnected datasets — without being trained separately for each new task.
One model, any timeline
A time-series foundation model works differently from the usual forecasting tools most teams build. Traditionally, a company builds one model for demand, another for machine failures and yet another for energy use, each fitted to its own history. Yug-1 aims to replace that patchwork with a single checkpoint: give it a historical context and a forecast horizon, and it forecasts a series it has never seen before, with no task-specific training and no per-series tuning.
The company says Yug-1 returns a predictive distribution, summarised by quantile levels from 0.1 to 0.9, rather than a single point estimate. In practice, that means a planner does not just get one number for next week — they get a range of likely outcomes, so they can prepare for the optimistic case and the difficult one at the same time. The model supports univariate forecasting and covariate-informed forecasting within a single architecture, so related signals can be read together: a target series plus the channels that move with it.
Under the hood, Yug-1 is a 271.8-million-parameter model, with a model dimension of 960, 12 decoder layers and 8 attention heads. It takes a maximum context length of 2,048 points, uses a patch-based decoder that emits a whole block of future points per forward pass, and runs in float32 precision. That patch-based design is part of what makes it fast: there is no per-step generation loop slowing down a long forecast.
Small by design, built for real use
At 271.8 million parameters, Yug-1 is deliberately compact next to the headline chatbots that dominate AI news. The company built it to be cheap to run and easy to try: developers can install it with pip install yug, check their machine with yug-check, and start forecasting from the model card on Hugging Face and the code on GitHub.
More Read
Birla AI Labs describes Yug-1 as Asia’s first time-series foundation model built for real-world use. Its frontier research team took the model from training to deployment in under six months, and presented the work at NeurIPS 2025, ICML 2026 and the World Summit AI, as well as at university campuses in India, Europe and the UK. On the GiftEval benchmarking framework, the company says Yug-1 outperformed other first-generation time-series foundation models and performed competitively against newer-generation models.
The training data behind the model combines a subset of the Chronos datasets and a subset of the GIFT-Eval pretraining data with proprietary synthetic data. The model weights are available under a CC BY-NC 4.0 licence, while the code is licensed separately under Apache 2.0 — an open setup that lets researchers and developers test the model, stretch it and build with it, and tell the team what they find.
Where a forecaster like this earns its keep
The uses for a model like Yug-1 touch everyday life more directly than they first appear. A retailer forecasting demand can stock the right goods before a festival rush instead of running out or over-ordering. A factory watching equipment performance can spot an unusual reading before a machine stops, saving a shift’s output. An energy operator forecasting load can plan generation more cleanly as solar and wind grow. A finance team tracking signals across markets can see how different variables move together, rather than studying each one in isolation.
That breadth is the point. Demand and supply, equipment performance, financial signals and infrastructure use all move through time, and the company’s founder and chairperson, Ananya Birla, puts it simply: much of the data that runs businesses and economies moves through time, and understanding how those patterns evolve is where some of the most consequential decisions get made. Yug-1 is built for exactly that work — forecasting outcomes, finding patterns and detecting anomalies across complex, interconnected variables.
For India, the release also carries a wider significance. The country has produced world-class language models and application software; a home-grown foundation model for time series adds a different layer to that stack — the layer that helps factories, grids, shops and hospitals plan ahead with numbers they can trust. Built in India and shared openly with researchers around the world, Yug-1 puts an Indian name alongside the global time-series models from Google and Amazon, and gives Indian teams a model trained with their own use cases in mind.
Yug-1 is version 0.1.1 of a longer journey: the team plans to present the work at NeurIPS 2026 and to bring it into the hands of tech and business leaders at campuses and labs globally in the coming months. For now, the invitation is open — try it, stretch it, build with it. The future, after all, is a time series, and India now has a model built to read it.
Keep reading: Built to Be Worn: Inside QWR, the Pune Company Making India’s Own VR Headsets and AI Glasses · On Wheels, Not Legs: Inside Addverb’s Elixis-W, India’s First Wheeled Humanoid Robot · More Industrial Automation & Technology stories
Like stories like this? Add Exceptional Technology as a preferred source on Google to see more of our reporting in your search results and Discover feed.
