Why we are featuring it
Time-series forecasting with a pretrained general model.
TimesFM applies a decoder-only foundation model to time-series data. Version 2.5 reduces parameters, expands context, and adds continuous quantile forecasting and fine-tuning support.
What is inside
- TimesFM 2.5 uses 200M parameters and up to 16k context.
- Quantile forecasts up to a 1k horizon.
- PyTorch and Flax model versions.
- Examples with XReg, Transformers, PEFT, and LoRA.