Hirondel-7M: zero-shot probabilistic time-series forecasting

Hirondel — from French hirondelle (swallow), the bird said to tell the weather.

Hirondel is a 7M-parameter time-series forecasting model for zero-shot probabilistic forecasting: 9 quantiles from 0.1 to 0.9, any horizon, CPU-friendly. Code and weights under Apache-2.0: https://github.com/tncrd/hirondel.

One of the most accurate small time-series foundation models. At release (October 2026), it was the most accurate on GIFT-Eval, fev-bench and TIME among models under 10M parameters with public weights and a permissive license (official leaderboards, read on 2026-10-23; how this was checked: showcase/README.md).

Results

These numbers are computed by us and not yet listed on the leaderboards. Added to the fev-bench leaderboard, they would rank 7 of 33 (SQL) and 6 of 33 (MASE) among the models the site shows by default (2026-10-23).

Univariate setting: every model forecasts each series alone, without covariates (competitors run by us in forced univariate mode with their public packages; TiRex, TimesFM-2.5, FlowState-r1.1 and Moirai-2.0 univariate by design). Pareto fronts over models with a permissive license, the others shown hollow.

Industrial data: IND-Eval

Five industrial sources of IND-Eval, none in the training data (top row), then the nine of the paper's aggregate (bottom row). Hirondel: 0.839 / 0.536 (nMASE / nCRPS) on the five. With every model forecasting each channel alone (circles, our runs with the same harness), Hirondel is ahead of Chronos-2, TimesFM-2.5 and TiRex, level with Timestral-5M and Toto-2.0-22m, and behind Timestral-28M. Forecasting the channels jointly with PRONTO's covariates (crosses) moves Chronos-2 from 0.895 to 0.822. Tables per source, protocol and model versions: README, IND-Eval.

Official leaderboards, read on 2026-10-23 (some models use covariates or joint multivariate forecasting there). Sources: showcase/data/ in the code repository.

Usage

import numpy as np
from hirondel import HirondelForecaster

forecaster = HirondelForecaster.from_pretrained("Tncrd/Hirondel-7M")
t = np.arange(1000)
context = np.sin(2 * np.pi * t / 24) + 0.1 * np.random.default_rng(0).normal(size=t.size)
forecast = forecaster.forecast(context, prediction_length=48)
print(forecast.quantiles.shape)   # (1, 48, 9)

Install with pip install "hirondel @ git+https://github.com/tncrd/hirondel@v1.1.0" (* pip install hirondel: PyPI release in preparation). Inference details, tail levels, hardware and numerical parity: see the code repository's README (https://github.com/tncrd/hirondel).

Known-future covariates (v1.1)

Inputs known in advance over the horizon (calendar, promotions, prices, set-points) can be passed with future_covariates, without retraining: per series, a ridge regression of the target on the covariates over the context, the model forecasts the rest, and their effect over the horizon is added back to every quantile. A gate (the default) keeps the correction only when it beats the plain forecast by 2 % on the end of the context.

day = np.arange(1000 + 48) % 7                          # covariate over context + horizon
context = context + 2.0 * (day[:1000] >= 5)             # weekends are higher
forecast = forecaster.forecast(context, 48, future_covariates={"weekend": day >= 5})
print(forecast.covariate_used)                          # [ True]: the gate kept the covariate

On the 42 fev-bench tasks with covariates, the skill score (SQL) goes from 41.5 to 44.5 (Chronos-2, which reads covariates natively: 47.0). Per task, the error changes by a factor ×0.949 on average (geometric mean), from ×0.51 (epf_de) to ×1.11 (walmart), the worst case; 14 tasks improve, 12 get worse. On PRONTO (IND-Eval), nMASE goes from 0.956 to 0.870 and nCRPS from 0.853 to 0.701; over the five industrial sources (the other four have no covariates and are unchanged), nCRPS goes from 0.536 to 0.515. The 12 fev-bench tasks whose covariates are known only in the past are unchanged. Over all 100 fev-bench tasks (the others unchanged), the skill score goes from 44.9 to 46.1 (Chronos-2: 47.3).

Linear and instantaneous effect; covariates known only in the past and static ones are not used. Without covariates the output is the univariate one, bit for bit; covariate_mode="off" turns them off. Details: the code repository's README, section "Known-future covariates".

Training data

Real data under licenses that allow this release, plus synthetic data. Every source, its license, its attribution and the changes made are listed in DATA.md; NOTICE gathers the required attributions. No dataset of GIFT-Eval's test split, fev-bench or TIME is in the training data.

Limitations

• Univariate: no cross-series information; known-future covariates only through a training-free per-series ridge (v1.1, future_covariates; linear, instantaneous).
• Context: 2 048 points at full resolution plus a coarse reading of up to 8 192 points.
• Tail levels 0.01 / 0.05 / 0.95 / 0.99 are extrapolated, not learned.

License

Apache-2.0 for code and weights. NOTICE must be kept when redistributing.

Citation

@misc{hirondel2026,
  title        = {Hirondel-7M: zero-shot probabilistic time-series forecasting},
  author       = {{The Hirondel Author (Tncrd)}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/Tncrd/Hirondel-7M}}
}