Our Take
Darts solved a real friction point—foundation models shipping as isolated packages—but this is plumbing, not progress; the forecasting accuracy claims belong to the underlying models, not the wrapper.
Why it matters
Foundation models for time series forecasting have fragmented rapidly across packages and interfaces. Teams evaluating or deploying multiple models face integration overhead; this removes that friction and lets practitioners test which model fits their pipeline without rewiring.
Do this week
Time series teams: benchmark your current forecasting pipeline against Chronos-2 and TimesFM 2.5 using the new Darts interfaces before next sprint planning, so you can decide if zero-shot foundation models cut your retraining costs.
Darts wraps four foundation models under one API
Darts, the open-source Python library for time series analysis originally released in 2020, has added a unified FoundationModel class that integrates Chronos-2, TimesFM 2.5, TiRex, and PatchTST-FM. The update standardizes interfaces across models that previously shipped as isolated packages with fragmented tooling.
The new implementation sits within the existing Darts ecosystem, meaning practitioners can swap foundation models by name change alone in existing pipelines. New pipelines can use the models for zero-shot forecasting, fine-tuned forecasting, uncertainty estimation, and backtesting, all combined with Darts' native data processing and evaluation tooling.
The work addresses a documented friction point: foundation models for time series, while individually published with accuracy claims, lacked common interfaces and limited interoperability with standard workflows. This made joint evaluation and production integration difficult without custom glue code.
Integration overhead was real; accuracy claims remain model-specific
The value here is operational. Teams already using Darts can now test foundation models without abandoning their pipeline. Teams evaluating multiple models no longer need to build separate adapters for each one. That eliminates a non-trivial engineering tax.
What does not change: the underlying accuracy of each model is independent of the wrapper. Chronos-2, TimesFM 2.5, TiRex, and PatchTST-FM made their own claims about zero-shot performance before this integration. Darts' contribution is orchestration and standardization, not model improvement.
For practitioners already invested in Darts, this is low-friction optionality. For those not yet using the library, this is a reason to consider it as the base layer for time series work that might involve foundation models.
Test which model fits your data before committing
Start with your current Darts pipeline and your dataset. Use the FoundationModel interface to run zero-shot forecasts with each of the four models. Compare against your existing custom or off-the-shelf baseline on the same test set and evaluation metrics you already trust.
Do not assume zero-shot works for your domain; validate on your data first. If one model outperforms others, check whether fine-tuning on your labeled data improves further, and whether the added accuracy justifies the engineering cost. Log the results and pin the winning model version.
If you are not yet using Darts, audit whether consolidating your time series tooling around it makes sense. Darts now covers model variety (foundation models plus classical methods) in one ecosystem, which can reduce operational surface area if you are today juggling multiple libraries.