32 models · ranked by HuggingFace downloads
Chronos-2 is Amazon's second-generation pretrained foundation model for zero-shot time-series forecasting. It frames forecasting as a language modeling problem over quantized time-series tokens using a T5 encoder-decoder architecture, enabling it to forecast across diverse domains without per-dataset training. Released under Apache 2.0.
Chronos-Bolt-Small is a small time-series foundation model from AutoGluon, using a T5-based encoder-decoder architecture for zero-shot forecasting. The 'Bolt' variant improves over original Chronos through training and architectural refinements for better speed and accuracy. Apache 2.0 licensed and part of the AutoGluon time-series forecasting ecosystem.
AutoGluon's distribution of Amazon's Chronos-2 time-series foundation model, packaged for use within the AutoGluon machine learning framework. The model uses a T5 encoder-decoder over quantized time-series tokens for zero-shot forecasting across diverse domains. AutoGluon wraps it in a high-level API for automated time-series modeling pipelines.
Chronos-Bolt-tiny is Amazon's lightweight version of the Chronos forecasting architecture, distilled for faster inference. It trades forecast accuracy for lower latency, making it suitable for high-frequency or resource-constrained forecasting scenarios.
Chronos-2-small is Amazon's pre-trained time series forecasting model based on a language model architecture, translating time series into token sequences and generating probabilistic forecasts. Small variant designed for fast inference.
A custom tokenizer base model from the Kronos project, providing a vocabulary and tokenization scheme for research into novel tokenization strategies. Published as a standalone artifact for integration with custom training pipelines.
Chronos-Bolt-base is Amazon's distilled time-series foundation model, faster than the original Chronos while retaining strong zero-shot forecasting accuracy. Built on a T5-style encoder-decoder trained on a large corpus of diverse time-series datasets, it supports probabilistic output distributions. Bolt variants were specifically optimized for production inference latency.
chronos-bolt-small models temporal dependencies in sequential numerical data to produce multi-step predictions.
chronos-bolt-tiny models temporal dependencies in sequential numerical data to produce multi-step predictions.
chronos-t5-base is Amazon's Chronos T5-base, a zero-shot time-series forecasting foundation model that quantizes real-valued series into token sequences and applies language model pre-training on synthetic and real-world datasets. The base variant balances inference speed with forecast quality for univariate series. It requires no per-dataset fine-tuning to generate probabilistic forecasts.
Chronos-Bolt-Base is the base-size variant of Amazon's improved Chronos forecasting model series, using a T5 encoder-decoder architecture. The Bolt series improves training efficiency over the original Chronos through revised architectural choices, achieving better forecast accuracy at equivalent model sizes. Apache 2.0 licensed.
chronos-t5-tiny predicts future values in time-series data using a T5 architecture conditioned on historical context windows.
Kronos-base predicts future values in time-series data using a transformer architecture conditioned on historical context windows.
Kronos-small predicts future values in time-series data using a transformer architecture conditioned on historical context windows.
chronos-bolt-mini models temporal dependencies in sequential numerical data to produce multi-step predictions.
chronos-t5-small predicts future values in time-series data using a T5 architecture conditioned on historical context windows.
MOMENT-1-small is a pre-trained time-series foundation model from CMU's Auton Lab, trained on the Timeseries-PILE collection to support multiple tasks including forecasting, classification, anomaly detection, and imputation without task-specific architecture changes. The small variant reduces parameter count relative to larger MOMENT checkpoints while retaining the T5-encoder-based backbone described in arXiv:2402.03885. MIT license and Transformers compatibility make it straightforward to integrate.
moirai-2.0-R-small predicts future values in time-series data using a transformer architecture conditioned on historical context windows.
chronos-t5-large performs multivariate or univariate time-series forecasting. It encodes temporal patterns and projects them into configurable forecast horizons.
Kronos-mini models temporal dependencies in sequential numerical data to produce multi-step predictions.
Kronos-Tokenizer-2k is a time-series forecasting model designed for financial candlestick (K-line) data, with a context window of 2,000 time steps. It applies a tokenization-based approach to OHLCV sequences, as described in arxiv:2508.02739. The MIT license and PyTorch/safetensors packaging make it straightforward to integrate into quantitative research workflows.
moirai-1.0-R-base is Salesforce's MOIRAI universal forecasting model, a transformer trained across a diverse mixture of time-series domains using the UNI2TS framework. It supports variable frequency (hourly, daily, weekly, etc.) and multivariate series with patch-based tokenization. The base variant is suitable for general-purpose zero-shot forecasting evaluation.
Granite-TTM-R2 (TinyTimeMixer r2) is IBM's compact foundation model for time series forecasting, pre-trained on diverse time series datasets. It uses a lightweight mixer architecture and supports zero-shot or fine-tuned forecasting with minimal compute. The R2 revision improves over the original TTM release with extended pre-training.
TiRex is NX-AI's zero-shot time-series forecasting model built on the TiDE/TiRex architecture with ONNX export for hardware-agnostic inference. It is pre-trained on large public time-series corpora and produces probabilistic point forecasts without dataset-specific fine-tuning. ONNX format enables deployment in environments without PyTorch.
IBM Granite's TinyTimeMixer R3 (ttm-r3) is a compact pre-trained foundation model for time series forecasting. It uses a mixing-based architecture (not a transformer) to handle univariate and multivariate forecasting tasks across diverse domains without task-specific training.
granite-timeseries-ttm-r1 models temporal dependencies in sequential numerical data to produce multi-step predictions.
TimeMoE-200M performs multivariate or univariate time-series forecasting. It encodes temporal patterns and projects them into configurable forecast horizons.
TimeMoE-50M is a 50M-parameter Mixture-of-Experts model designed for zero-shot univariate time-series forecasting. It frames forecasting as next-token prediction over continuous values, using sparse expert routing to handle diverse time-series distributions without dataset-specific fine-tuning. The model targets multi-horizon predictions across domains like energy, finance, and weather.
Moirai-MoE is Salesforce's Mixture-of-Experts time-series foundation model for zero-shot forecasting across diverse domains. The small variant balances capability and inference cost for the time-series forecasting task.
moirai-1.1-R-large performs multivariate or univariate time-series forecasting. It encodes temporal patterns and projects them into configurable forecast horizons.
sundial-base-128m predicts future values in time-series data using a Chronos architecture conditioned on historical context windows.
chronos-2-synth is an open-weight checkpoint for time-series forecasting, distributed on the HuggingFace Hub. The Apache 2.0 license keeps chronos-2-synth unrestricted for commercial reuse. chronos-2-synth is community-maintained, so track upstream changes and pin a known-good revision.