IBM opens a 385M time-series forecaster

IBM released Granite Time Series PatchTST-FM-r2, a roughly 385-million-parameter zero-shot forecaster with 8,192-step context and 99 quantile outputs. Its model announcement says it ranks second among replicable zero-shot models on GIFT-Eval and is dual-licensed under Apache 2.0 and OpenMDW 1.0; the model card notes IBM’s result is still in a pending benchmark pull request. Teams evaluating forecast models can test it under commercial-friendly terms, but should reproduce the leaderboard result on their own demand or telemetry data, as Spark X25’s token-cost break-even shows, licensing and unit economics are separate decisions.