Timer-S1
onlinebytedance-research/Timer-S18.3B total / 0.75B active params | 12K context | $0.00025 per forecast | Apache-2.0
Timer-S1 is ByteDance's billion-scale sparse MoE time-series foundation model and the high-capacity entry in the Timer line. With 8.3B total parameters but only 0.75B active per token, it is positioned as a top-tier zero-shot forecaster, reaching state-of-the-art MASE and CRPS on the GIFT-Eval leaderboard.
Architecturally it is a decoder-only sparse Mixture-of-Experts transformer built around Serial-Token Prediction (STP) and an 11.5K-token context window, and it natively emits 9-quantile probabilistic forecasts rather than a bare point estimate. It is pretrained on TimeBench, a curated corpus of one trillion time points spanning diverse domains, with data augmentation applied to mitigate predictive bias, and it is released under Apache-2.0.
On TSFM.ai reach for Timer-S1 when you want the strongest available Timer-family forecaster with long-context, probabilistic output and can afford a billion-scale model. Drop to THUML's 84M Timer when you only need lightweight zero-shot point forecasting from the original Timer architecture and the much larger MoE checkpoint is more than the workload requires.
Model Classification
Family
Timer-S1
Type
time series foundation model
Pretrained time-series model exposed on TSFM.ai for zero-shot or few-shot forecasting workloads.
Resources
Training Data
TimeBench, a curated corpus with one trillion time points spanning diverse domains, with data augmentation to mitigate predictive bias.
Recommended For
- • Zero-shot univariate forecasting with causal autoregressive generation
- • Long-horizon prediction across heterogeneous time-series domains
Strengths
- • TimeAttention unifies variable-length and multi-resolution inputs
- • Strong zero-shot performance from 260B-point pretraining
Limitations
- • Smaller model family with fewer checkpoint size options than Chronos or Moirai
- • Causal-only architecture limits suitability for bidirectional tasks like imputation
- • Hosted Timer serving works best once you have at least one full 96-point patch of history; very short series are a poor fit
- • The current hosted checkpoint can flatten simple repeated seasonal toy probes more than newer specialist zero-shot models
Not Ideal For
- • Histories shorter than one full 96-point patch
- • Users who need strong seasonal continuation on very small repeated-pattern probes without tuning
Capabilities
Tags
Specifications
- Parameters
- 8.3B total / 0.75B active
- Architecture
- decoder-only sparse MoE transformer with Serial-Token Prediction (STP)
- Context length
- 11,520
- Max context
- 11,520
- Minimum history
- n/a
- Recommended history
- n/a
- Input step
- n/a
- Required target series
- 1
- Temperature
- Ignored
- Top P
- Ignored
- Max output
- 2,048
- Avg latency
- n/a
- Uptime
- n/a
- Plan limits
- 1,000 rpm free · 1,000,000 rpm with billing
- Accelerator
- A10G
- Regions
- Virginia, US
- License
- Apache-2.0
Pricing
- Per forecast
- $0.00025
Performance
- Average latency
- n/a
- Availability
- n/a
- Plan limits
- 1,000 rpm free · 1,000,000 rpm with billing