Yug-1 can forecast a range of possible outcomes for a previously unseen time series without training a separate model for that dataset, according to Birla AI Labs. The 271.8-million-parameter model is downloadable and supports additional input variables. Forecasting teams can install its Python package and test it against their own historical data.
The release targets problems such as demand planning, equipment monitoring, finance and infrastructure, according to The Economic Times’ launch report, published on October 19, 2026.
Before treating it as a production candidate, teams face two qualifications: its reported GIFT-Eval results are competitive without leading across the board, and the downloadable weights carry a noncommercial license. Being able to test Yug-1 doesn't establish its suitability or permission for a business deployment.
Yug-1 Forecasts Numeric Series, Not Language
Despite the foundation-model label, Yug-1 isn't an LLM for answering questions about spreadsheets. It accepts historical numeric observations and a forecast horizon, then produces predictions for future observations.
The Yug repository describes a single pretrained checkpoint that works without task-specific training or per-series tuning. That's the practical meaning of zero-shot here: a team can supply a new series without first fitting the model to that individual forecasting problem. This describes the workflow; it doesn't guarantee accuracy on every dataset.
The model supports up to 2,048 historical observations. Its published architecture has 12 decoder layers, a model dimension of 960 and eight attention heads. A patch-based decoder produces blocks of 64 future points per forward pass instead of generating one point at a time. Those details explain the decoding approach, though latency still needs to be measured on a team's hardware.
Outputs include nine quantile levels, from 0.1 through 0.9. The median provides a central forecast; the outer quantiles describe uncertainty. For demand planning, that lets a team examine a high-demand scenario alongside an expected sales figure.
Those bands need validation. An interval between the 0.1 and 0.9 quantiles nominally represents an 80% prediction interval. Teams should measure whether it actually captures roughly 80% of held-out observations.
Covariates Come With a Long-Horizon Warning
Yug-1 supports both univariate forecasting, using the target series alone, and forecasting with covariates: additional variables supplied alongside the target.
