Yug

Yug is a pretrained time-series foundation model developed by Birla AI Labs for zero-shot probabilistic forecasting. A single checkpoint forecasts a previously unseen series without task-specific training or per-series tuning: given a historical context and a forecast horizon, it returns a predictive distribution summarised by quantile levels.

• Model weights: birlaailabs/yug
• Package: pip install yug
• Issues and questions: GitHub Issues

Latest version: Yug 0.1.1

---

Update: October 2026

First public release of Yug.

Key highlights:

• Zero-shot forecasting. One pretrained checkpoint forecasts any evenly spaced numeric series, across domains and frequencies, without fitting.
• Probabilistic by default. Every forecast comes back as quantile levels, not a single line, so intervals are available at no extra cost.
• Fast direct multi-step decoding. A patch-based decoder emits a whole block of future points per forward pass; there is no per-step generation loop.
• Covariate support. Optional additional channels alongside the target.
• Drop-in benchmarking. A GluonTS predictor for existing evaluation harnesses.

---

Available models

Model	Parameters	Quantile levels	Context length
birlaailabs/yug	271.8M	0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9	2048

---

Install

From PyPI:

pip install yug

Optional extras:

pip install 'yug[pandas]'     # used in the example below
pip install 'yug[gluonts]'    # GluonTS predictor for benchmark harnesses

Check the machine can run it before downloading weights:

yug-check

Weights are cached locally on first use (~/.cache/huggingface/hub by default, or HF_HOME if set).

---

Code examples

1. Forecast one series

import numpy as np
from yug import YugPipeline

pipeline = YugPipeline.from_pretrained(
    "birlaailabs/yug", device_map="cuda"
)

# Synthetic signal: trend + 64-step season + 12-step season + noise
CONTEXT, HORIZON, PERIOD = 2000, 256, 64
rng = np.random.default_rng([20240902, 315595964, 0])
t = np.arange(CONTEXT + HORIZON)
trend = np.linspace(0.0, rng.uniform(50, 150), len(t))
phase1, amp1 = rng.uniform(0, 6), rng.uniform(10, 30)
phase2, amp2 = rng.uniform(0, 6), rng.uniform(2, 8)
signal = (
    trend                                                 # trend
    + amp1 * np.sin(2 * np.pi * t / PERIOD + phase1)      # 64-step season
    + amp2 * np.sin(2 * np.pi * t / 12 + phase2)          # 12-step season
    + rng.normal(0, 1.0, len(t))                          # noise
).astype(np.float32)

context, truth = signal[:CONTEXT], signal[CONTEXT:]

forecast = pipeline.predict(
    context=context,
    prediction_length=HORIZON,
    freq="D",            # a model input, not metadata: pandas offset alias
    num_samples=30,      # trajectories drawn to estimate the quantiles
    seed=0,              # makes the call reproducible
)

print(forecast.median)                # (256,)   median (point) forecast
print(forecast.quantile(0.9))         # (256,)   upper quantile
print(forecast.interval())            # {"lower": (256,), "upper": (256,)}
print(forecast.to_dataframe().head())

Note

• num_samples: trajectories drawn to estimate the quantile band. More samples give smoother, more accurate quantiles but run slower; the default is 100. On CPU, a lower value (e.g. 30) is usually a good trade-off.

2. Several series at once

Lengths and frequencies may differ.

forecast = pipeline.predict(
    [series_a, series_b, series_c],
    prediction_length=48,
    freq=["D", "D", "H"],
    item_ids=["a", "b", "c"],
)
forecast.values          # (3, n_quantiles, 48)

3. With covariates

Stack channels into a 2-D array, target first, and pass it as a single series:

multivariate = np.stack([target, covariate_1, covariate_2])
forecast = pipeline.predict([multivariate], prediction_length=48)

4. In a GluonTS benchmark

from yug.gluonts_adapter import YugPredictor

predictor = YugPredictor.from_pretrained(
    "birlaailabs/yug", freq="H", prediction_length=48
)
forecasts = list(predictor.predict(dataset))

---

Examples

• notebooks/quickstart.ipynb: forecast, plot and score a series end to end. Self-contained; generates its own data.
• notebooks/gift_eval.ipynb: run the GIFT-Eval benchmark to reproduce the published scores.

---

Citation

@software{yug_2026,
  title  = {Yug: a foundation model for zero-shot probabilistic time-series forecasting},
  author = {Aaditya Jain* and Debdeep Sanyal* and Aaryan Nagpal and Dhruv Kumar and Murari Mandal and Saurabh Deshpande},
  year   = {2026},
  organization = {Birla AI Labs},
  url    = {https://github.com/birla-ai-labs/yug},
  license = {Apache-2.0}
}

Note: Aaditya Jain and Debdeep Sanyal contributed equally.

License

Code in this repository is licensed under the Apache License 2.0. Model weights on Hugging Face are licensed separately under a noncommercial license, see the model card for terms.