Request early access to ChakraTS-OS-10M

ChakraTS-OS-10M is in early access. We review requests by hand, usually within a day.

Log in or Sign Up to review the conditions and access this model content.

ChakraTS-OS-10M

ChakraTS-OS-10M is an open-weight time-series forecasting foundation model from YHat Labs. With 10.6M parameters it is frontier for its size, more accurate than most open models up to 100 times larger. It natively supports covariates (past and known-future, numeric or categorical) and multivariate forecasting, with a 32k context.

Library ยท PyPI ยท Usage guide ยท License ยท Hosted API

ChakraTS-OS-10M forecasting a week of hourly electricity demand on the Texas Gulf Coast, next to what happened

A week of hourly electricity demand on the Texas Gulf Coast, forecast from two weeks of history, with no training. Data: ERCOT.

It is pretrained on synthetic and real data, so it forecasts your series zero-shot: give it the past values of any series, such as daily sales, hourly electricity use or weekly sign-ups, and it predicts the next values with a range of likely outcomes around each one. It can use what you already know about the coming days, such as planned prices, promotions, holidays or a weather forecast. There is no model of your own to train, tune or maintain.

  • Ranges, not just numbers: calibrated quantiles or intervals at any level, for one series or thousands.
  • Small and fast: under 0.1 s for a 48-step forecast of one series on a laptop CPU; NVIDIA and Apple GPUs too.

Quick start

pip install yhatmodels
import numpy as np
from yhatmodels import ChakraTS

model = ChakraTS.from_pretrained("ChakraTS-OS-10M")

history = 10 + np.sin(np.arange(500) / 24 * 2 * np.pi)  # any series: sales, demand, sensor readings, ...
q = model.forecast(history, h=48, level=[80, 95])         # the median and the 80% and 95% intervals, 48 steps ahead

During early access, request access on this page, then sign in with hf auth login once you're approved.

The first download asks you to accept the license once with a free yhatlabs.com account: your browser opens, you sign in or create an account and accept, and the download starts. After that the model runs entirely on your computer, and your data never leaves it. In CI or on a server, set YHAT_MODELS_TOKEN (copy it from yhatlabs.com/models/license).

Use what you know about the future

Name the columns you know ahead of time as known, and put their future values in the rows after your last observation:

# df: one row per hour with ds, y (solar output), sunlight and clouds;
# the last 48 rows hold tomorrow's weather forecast, with y empty
fc = model.forecast(df, h=48, known=["sunlight", "clouds"])

Two days of solar output forecast from history alone and with sunlight and cloud cover

Two days of solar output. From history alone, the model expects another sunny day; with sunlight and cloud cover it sees the cloudy one coming. Data: Renewable Energy and Weather Conditions (CC0).

The library also backtests on your own history, plots forecasts and calibration, fine-tunes on your data and runs from the command line; see the library.

Results

fev-bench scores forecasting models on 100 tasks from many domains. Its skill score says how much more accurate a model is than a simple seasonal forecast; higher is better. ChakraTS-OS-10M is the most accurate open model under 100M parameters, and more accurate than most open models up to 100 times its size.

fev-bench skill score against model size for open forecasting models

It handles covariates well: on the 42 tasks that come with them, it keeps its accuracy, while most models lose some:

Model Parameters All 100 tasks 42 tasks with covariates
ci-proto 167M 51.5 53.0
TimeH2O-v1 109M 50.8 51.4
Timestral-209M 209M 50.4 50.7
ChakraTS-OS-10M 10.6M 49.9 50.1
t0-gamma 256M 49.4 49.3
TimesFM-3 331M 49.0 49.0
Simons-297M 297M 48.4 47.3
Chronos-2 119M 47.3 47.0
Timestral-5M 5.2M 46.2 44.5
TimeH2O-v1-tiny 5.8M 45.5 41.5
Toto-2.0-2.5B 2.5B 44.4 38.8

fev-bench skill score (%), as of November 2026, computed the way the fev-bench leaderboard does from the result files in the fev repository; sizes from the models' Hugging Face pages.

On the two other public forecasting benchmarks:

Benchmark MASE CRPS
GIFT-Eval (97 configurations) 0.674 0.461
TIME (98 configurations) 0.642 0.537

Geometric means relative to Seasonal Naive; lower is better. All results use the default pro preset and each benchmark's own evaluation code.

Model details

Developer YHat Labs
Parameters 10.6M
Architecture decoder-only transformer over patches of 16 steps, attending across time and across a series' variables (targets and covariates)
History up to 32,768 past steps per series
Horizon any; long horizons are decoded 768 steps per pass
Output 21 quantile levels from 0.01 to 0.99; any level in between is interpolated
Covariates past and known-future, numeric or categorical
Presets pro (default): reads the series at several patch sizes and averages; fast: one patch size, several times faster
Training data synthetic series and public real-world time-series data
Precision bf16 on NVIDIA GPUs, fp32 elsewhere
Files model.safetensors (weights, sha256 in config.json), config.json, LICENSE.md, NOTICE

Intended use

  • Research on forecasting and time-series models, including benchmarking against other models.
  • Personal projects and data science competitions.
  • Internal evaluation by an organization, for up to 90 days, before adopting the model or a YHat Labs product.

Production use, client work, business decisions and hosted services need a commercial license or ChakraTS, our flagship model, through the YHat Labs API.

Limitations

  • A small model: holidays and other rare calendar effects, long multivariate series and abrupt level shifts are harder for it than for larger models.
  • It returns quantiles, not sample paths, and has no per-series attributes (static covariates).
  • Series without a regular time index are not supported; very short histories give wide intervals.
  • Forecasts are estimates; check them on your own history (model.evaluate) before relying on them.

License

The weights are released under the Chakra Non-Commercial Model License, Version 1.0 (also in LICENSE.md); the yhatmodels code is Apache-2.0. Anything you distribute or demo publicly that uses the model must say "Built with ChakraTS-OS by YHat Labs" and link to the license. For commercial use, use the hosted API or write to licensing@yhatlabs.com.

Citation

@misc{yhatlabs2026chakratsos,
  title        = {{ChakraTS-OS}: open-weight time-series forecasting models},
  author       = {{YHat Labs}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/yhatlabs/ChakraTS-OS-10M}},
  url          = {https://huggingface.co/yhatlabs/ChakraTS-OS-10M}
}

Contact

founders@yhatlabs.com ยท yhatlabs.com

Downloads last month
24
Safetensors
Model size
10.6M params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Evaluation results