IBM 发布约 3.85 亿参数的 Granite 时间序列模型 r2
IBM releases the roughly 385-million-parameter Granite time-series r2 model
Granite Time Series PatchTST-FM-r2 支持零样本预测、概率预测与缺失值填补,并公开权重、推理流程及复现代码。IBM 报告其在 9 月 8 日的 GIFT-Eval 可复现零样本比较中排名第二,并强调商业友好的许可。排名来自团队所列筛选条件与基准,不能直接外推到任意业务数据。
Granite Time Series PatchTST-FM-r2 supports zero-shot and probabilistic forecasting plus missing-value imputation, with weights, inference pipelines, and reproduction code released. IBM reports second place in its September 8 comparison of replicable zero-shot GIFT-Eval models and highlights commercially friendly licensing. The ranking depends on the stated filters and benchmark; it does not establish performance on arbitrary business data.