Chronicle: A Multimodal Foundation Model for Joint Language and Time Series Understanding

In three linesChronicle is a 324M-parameter multimodal foundation model trained from scratch on natural language and time series in a unified architecture. Both modalities share the same transformer blocks and attention mechanisms. It matches Gemma-3-270M on 19 NLU tasks, sets new benchmarks on 24 UCR/UEA datasets, and outperforms supervised fusion baselines on Time-MMD.

Summary generated by Claude — human-verified