Two years ago, "add forecasting to the product" meant hiring someone who knew ARIMA from a hole in the ground, feature-engineering a training set, and babysitting a model that drifted the moment your traffic pattern changed. In mid-2026 it means picking a checkpoint. Time-series foundation models forecast zero-shot — you hand one your history, ask for the next N steps, and it answers, no training run at all. It's the same jump TabPFN brought to tabular data: the hard part moved from building the model to choosing it.

And choosing is now the whole game, because the field learned a lesson that runs against the LLM instinct: bigger stopped winning.

The one thing to internalize: match the data's shape#

Google's TimesFM 2.5 shipped smaller than the model before it — around 200M parameters — and got better. Salesforce's Moirai-2 threw out its encoder for a leaner decoder-only design. Datadog's Toto-2 is 7x more parameter-efficient than Toto-1 at matching quality. Nobody is scaling a general forecaster to a trillion parameters and calling it done, because time series aren't language: a model tuned to the shape of your data beats a bigger generalist that isn't.

So the useful question isn't "which model tops the leaderboard." It's "what does my data look like?"

A mismatched giant loses to a fitted smaller one. The skill that used to be feature engineering is now model selection — and it's cheaper, faster, and more transferable.

What it looks like to ship one#

The reason this matters to a solo founder is the code is anticlimactic. For Chronos, AutoGluon runs the whole thing zero-shot:

from autogluon.timeseries import TimeSeriesPredictor

# no training data split, no hyperparameters — a foundation model preset
predictor = TimeSeriesPredictor(prediction_length=24).fit(
    history,                      # your past series, long-format
    presets="chronos",            # zero-shot: uses the pretrained model
    time_limit=60,
)
forecast = predictor.predict(history)   # quantile forecast, next 24 steps

Toto-2 and Moirai-2 ship open weights you load straight from Hugging Face and call the same way. The engineering effort that used to build the model now goes into two things worth doing well: picking the model that matches your data shape, and always backtesting against a dumb baseline. Which is the honest caveat —

When not to reach for one#

A foundation model is not automatically the answer. If your series is short, cleanly seasonal, and stable, a classic ETS/ARIMA baseline or a gradient-boosted model on lag features can match a TSFM for a fraction of the compute — the same "measure before you reach for the big tool" discipline that decides tabular problems. TSFMs earn their cost on the hard cases: cold-start series with no history to train on, hundreds of series you'd never hand-tune individually, and messy multivariate data where feature engineering doesn't scale.

The headline is still the good news. Forecasting used to be a project. Now it's a decision — and the whole decision is: look at the shape of your data, then pick the model built for that shape. Get that right and a 200M-parameter model you deployed on Tuesday will beat the leaderboard king you didn't need.