ModelsMedium Impact·Wednesday, September 9, 2026

IBM releases Granite Time Series PatchTST-FM-r2 with commercial-friendly license

IBM's new 385M‑parameter PatchTST-FM-r2 model tops zero‑shot time‑series forecasts while offering Apache‑2.0 licensing.

What happened

IBM announced the Granite Time Series PatchTST-FM-r2 model, a 385M‑parameter foundation model that achieves the best zero‑shot performance among permissively licensed models on the GIFT‑Eval benchmark. The model, code, and weights are released on Hugging Face under Apache‑2.0 and OpenMDW‑1.0 licenses and can be integrated with streaming pipelines via Confluent Cloud.

Why it matters to you

personalized

Implement zero‑shot forecasts by loading the model from Hugging Face and calling its inference API.

What to do about it

Generate a 24‑hour demand forecast for a CSV of hourly sensor readings using the provided HF pipeline.

Put it to workHugging Face Transformers · 5 min

A numeric array of 24 forecasted values matching the input series length.

Suggested exercise · not independently tested. Free access suggested; check current availability.

Before you start: Python 3.9+ · Hugging Face account (free)

1
Create a free Hugging Face account and install the transformers and datasets libraries.
2
Download the PatchTST-FM-r2 model with `from transformers import AutoModelForSeq2SeqLM, AutoTokenizer`.
3
Load your CSV, format the series as a list, and call `model.generate` with the desired horizon.
4
Print or save the forecast array.

Check your result: The output array length equals the requested horizon and values are numeric.

Community

Comments