Pre-trained on 100B points, but you’ll need an API key

What it does TimeGPT-1 is a generative transformer trained exclusively on time-series data—over 100 billion points across finance, energy, IoT, and retail—to deliver zero-shot forecasts and anomaly detection. The repository is a Python SDK and thin client that sends your pandas DataFrames to a hosted API and returns predictions, intervals, or anomaly flags. You can also fine-tune the model on your own data, add exogenous variables, and run it inside Snowflake via stored procedures.

The interesting bit Unlike LLM-based hybrids, TimeGPT was built from scratch as an encoder-decoder transformer for temporal patterns, not language. The README claims it ranks in the top three across various frequencies when tested zero-shot on more than 300,000 unique series, with an average GPU inference latency of 0.6 milliseconds per series—roughly on par with a seasonal naive baseline.

Key highlights

• Zero-shot forecasting and anomaly detection without local model training
• Support for exogenous variables, multiple series, irregular timestamps, and custom loss functions during fine-tuning
• Snowflake native deployment so data never leaves the warehouse
• Prediction intervals and built-in cross-validation
• Not based on an existing LLM; trained independently on time-series data only

Caveats

• The model is proprietary and accessed via API; you need a key from Nixtla’s platform (free trial available)
• Some advertised integrations, such as Azure Studio, are marked “coming soon”
• The README’s performance claims are internal benchmarks; independent verification isn’t shown

Verdict Worth a look if you want low-code forecasting without managing GPU infrastructure, but skip it if you need open weights, an air-gapped environment, or full transparency into the model architecture.