TiRex
v1.0.0onlineNX-AI/TiRex~300M params | 16K context | $0.2400 input | $0.4800 output | Apache-2.0
Nixtla's large-scale forecasting transformer with native exogenous covariate support and approximately 300M parameters. Features a 16K context window and built-in handling of holidays, promotions, and other external regressors. Successor to the TimeGPT architecture with improved zero-shot transfer and fine-tuning capabilities.
Model Classification
Family
TiRex
Type
foundation-model
Pre-trained time series model intended for zero-shot or minimal-adaptation use.
Resources
Training Data
Proprietary multi-domain corpus spanning retail, energy, finance, and logistics — 50B+ observations with exogenous covariate pairs
Recommended For
- • Forecasting with promotions, holidays, or weather covariates
- • Risk-aware probabilistic forecasting with quantiles
Limitations
- • Evaluate with your own historical backtests before tier-1 production routing.
Capabilities
forecastingprobabilistic-forecastingcovariateslong-context
Tags
populartier-1launch
Specifications
- Parameters
- ~300M
- Architecture
- transformer with covariate encoder
- Context length
- 16,384
- Max output
- 2,048
- Avg latency
- 260ms
- Uptime
- 99.92%
- Rate limit
- 300 rpm
- Provider
- tsfm (us)
- GPU Class
- NVIDIA H100
- Regions
- us-central1
- License
- Apache-2.0
Timeline
- Added to platform
- Sep 28, 2025
- Last updated
- Jan 18, 2026
Pricing
- Input / 1M tokens
- $0.2400
- Output / 1M tokens
- $0.4800
Performance
- Average latency
- 260ms
- Availability
- 99.92%
- Rate limit
- 300 rpm