// the find
google-research/timesfm
TimesFM (Time Series Foundation Model) is a pretrained time-series foundation model developed by Google Research for time-series forecasting.
A pretrained decoder-only transformer from Google Research that does zero-shot time series forecasting — you give it historical values, it predicts future ones, no training on your data required. The 2.5 model is 200M params with 16k context and optional quantile heads. It ships in both PyTorch and Flax, and is already embedded in BigQuery ML and Google Sheets.
Zero-shot forecasting actually works out of the box — no dataset-specific training pipeline to maintain. The 2.5 architecture shrank from 500M to 200M parameters while extending context from 2k to 16k, which is a real engineering win, not just a version bump. Continuous quantile forecasting up to 1k horizon via a separate 30M head gives calibrated uncertainty estimates without retraining the base model. LoRA fine-tuning via HuggingFace PEFT is now documented with working examples, so domain adaptation is practical.
XReg covariate support is bolted on externally rather than being native to the model architecture — if you have rich feature sets, you're not getting the full benefit of a proper multivariate model. The PyPI version history is chaotic: 1.3.0 for old checkpoints, 2.0.0 for current, but v2.5 is the actual latest model — three version schemes that don't line up. No built-in data preprocessing pipeline; you're expected to hand it clean, aligned numpy arrays, which is most of the real work in time series. The benchmark comparisons in the repo are against older baselines and don't include recent strong competitors like Chronos or Moirai.