IBM's 385M Granite forecaster: zero-shot, and free to ship
The PatchTST-FM-r2 update leads GIFT-Eval's permissive field: probabilistic forecasts, no fine-tuning, one pip install, small enough to run on a CPU.
Copy markdown
#1 among the licenses you can actually use
On the GIFT-Eval leaderboard it lands #1 among replicable zero-shot models with permissive licensing, and #2 overall — geometric-mean CRPS 0.467, MASE 0.6846 — trailing only Google's TimesFM-3 while beating Chronos-2, Timer-S1 and Toto. The gap to the leader is a rounding error.
Apache 2.0 or OpenMDW — ship it in a product
It's dual-licensed under Apache 2.0 and the Linux Foundation's OpenMDW 1.0, so it drops straight into a commercial product with no sales call — the gap that sinks a lot of 'open' forecasters. And at ~385M parameters it's small enough to run on a CPU or a modest GPU.
Zero-shot, with error bars, in a few lines
No fine-tuning: `pip install "granite-tsfm>=0.3.9"`, call `PatchTSTFMForPrediction.from_pretrained()`, and forecast series the model has never seen. r2 adds 99-quantile probabilistic output — real uncertainty bands — plus missing-value imputation, over an 8,192-step context.
Build this weekend: forecasts for your own metrics
Point it at demand, energy load, traffic or app telemetry and get calibrated forecasts and anomaly signals without training a model per series. Want it on a live stream? There's a Confluent Cloud early-access path for Apache Flink.