Time series model architectures
Compare patching, tokenization, attention, mixing, and model heads without confusing architecture with deployment.
Patching and tokenization
Patching groups adjacent observations into model inputs; tokenization represents values through a discrete vocabulary. These choices affect sequence length, resolution, and prediction mechanics. Chronos-T5 and Chronos-Bolt illustrate different paths within one family, so family names alone do not identify the inference contract.
Backbones and outputs
Encoder-only, decoder-only, encoder-decoder, mixer, and recurrent designs organize computation differently. A model's task head determines whether it emits forecasts, quantiles, anomaly-related outputs, or representations. An exported encoder is not the same artifact as the complete forecast model.
Architecture is not an operating guarantee
Parameter count and backbone do not establish production latency, price, accuracy, or browser support. Device, precision, context length, batching, runtime, and graph operators affect execution. Evaluate the exact checkpoint and deployment contract alongside its research architecture.