time-series-forecasting
HF group: Tabular · Status: ❌ not built · candidate v2+
What it is
Numeric series → future values. Foundation models for time series have matured fast (2024-26) and are increasingly LLM-shaped: tokenize values, train transformer, generate forward.
Open-weight models
| Model | Params | Released | License | Quality | Notes |
|---|---|---|---|---|---|
| Chronos-Bolt (Amazon) | 9-205 M | 2024 | Apache-2.0 | T5-based; tokenises values | Most mature option; runs on CPU. |
| Chronos-2 | 200 M | 2025-10 | Apache-2.0 | Univariate + multivariate + covariate | 300+ forecasts/sec on a single GPU. |
| TimesFM (Google) | 200 M | 2024 | Apache-2.0 | Decoder-only, patch-based | Strong zero-shot; trained on 100 B time points. |
| Lag-Llama | ~150 M | 2024 | Apache-2.0 | Probabilistic forecasts | Llama-style decoder; full distributions. |
| Moirai (Salesforce) | 14-311 M | 2024 | Apache-2.0 | Universal forecaster | Strong. |
| Timer-XL (THUML) | various | 2025 | Apache-2.0 | Long-horizon | Newer entrant. |
Infrastructure required
Inference
- ❌ Encoder-only / encoder-decoder inference (sharing the rail with
text-to-embeddingsince these are encoder-only tokenizer+transformer stacks).
Input
- Numeric series (Float64 vector + timestamps).
- Optional covariates (multivariate forecasting).
Output
- ❌ Future values.
- ❌ Optional probability distribution (for uncertainty quantification).
Storage
- ❌ Weights cache.
- App-side: historical series persisted via
locara-storage.
Interaction (IPC + SDK)
- ❌
forecast.predict({ series, horizon })IPC.
Capabilities (manifest)
capabilities.models[]for the forecaster.
Gaps
Whole stack. Listed here so we don’t lose the thread — forecasting is increasingly part of consumer apps (energy, finance, health) and the open-weight options are now genuinely strong. Candidate v2+ when a reference app motivates it.
See also
text-to-embedding— same encoder-only inference shapeout-of-scope— sibling tabular tasks- Index:
../modalities-and-models-survey.md