Long-context, small-parameter design: this model family achieves state-of-the-art zero-shot forecasting with a 209M-parameter checkpoint that handles extremely long histories (up to 32,768 timesteps) and emits probabilistic forecasts (21 quantiles). The surprising insight is that careful architecture and training let a sub-300M model outperform many larger pretrained models on standard forecasting benchmarks.
Key Capabilities
- Long-context forecasting: accepts up to 32,768 historical timesteps (1,024 patches of 32 steps), preserving recent context and enabling long-horizon patterns to influence short-term forecasts.
- Probabilistic multivariate outputs: jointly forecasts multiple variates and returns 21 non-crossing quantiles per timestep, enabling uncertainty-aware decisions.
- Covariate-aware attention: supports past-only and known-future covariates (numeric and categorical, with categorical target-encoding), so calendar effects and external signals can be incorporated without leaking target information.
- Efficient, patch-based transformer: 24 layers, 768-dim embeddings, patching at 32 steps and block-by-block decoding for horizons beyond 1,024 steps; optimized to run on CPU or GPU with safetensors weights.
- Strong zero-shot performance: ranks first among pretrained/zero-shot models on GIFT-Eval, TIME and fev-bench, with competitive MASE and CRPS metrics.
Who it's for and trade-offs
Great fit if you need off-the-shelf, uncertainty-aware forecasts from long histories without task-specific fine-tuning — for exploratory forecasting, cross-domain baselines, or pipelines where CPU inference matters. It is also suitable when you want a compact foundation model (209M) that still matches or exceeds larger models on benchmark accuracy.
Look elsewhere if you require model interpretability at the feature level, extremely low-latency microsecond inference on-device (smaller family members or distilled variants may be faster), or if you need fully open permissive licensing for commercial redistribution beyond the model's Modified MIT constraints. Also, horizons beyond 1,024 steps are produced block-by-block using median feedback, which can affect multi-step joint consistency compared with models trained end-to-end for very long horizons.
