chronos-2-small

Technical Overview

Chronos‑2‑Small is the lightweight 28 million‑parameter variant of the Chronos‑2 family, a series of foundation models designed for universal time‑series forecasting. Built on a T5‑style encoder‑decoder architecture and released by AutoGluon, the model can ingest a wide variety of univariate and multivariate time‑series signals and generate accurate point forecasts as well as probabilistic intervals for horizons ranging from a few steps to several months.

Key features and capabilities include:

• Supports both univariate and multivariate inputs without the need for manual feature engineering.
• Provides probabilistic forecasts (e.g., quantiles) out‑of‑the‑box, enabling risk‑aware decision making.
• Handles irregular sampling and missing values via built‑in masking tokens.
• Optimized for fast inference on consumer‑grade GPUs thanks to its modest parameter count.
• Comes in .safetensors format, which is both memory‑efficient and secure.

Architecture highlights:

• Based on the T5 transformer encoder‑decoder, repurposed for sequential numeric data instead of text.
• 28 M trainable parameters, split roughly evenly between encoder and decoder stacks (6 layers each, 8 attention heads per layer).
• Positional encoding is adapted for time‑step intervals, allowing the model to understand irregular time gaps.
• Uses a continuous‑valued embedding layer that maps raw numeric observations to a high‑dimensional latent space.
• Output head predicts a mixture of Gaussians, which is then transformed into quantile forecasts.

Intended use cases:

• Retail demand forecasting (daily/weekly sales, inventory planning).
• Financial time‑series prediction (stock prices, volatility, macro‑economic indicators).
• Energy consumption and load forecasting for utilities.
• IoT sensor data analytics where devices produce irregular streams.
• Any domain that requires a fast, out‑of‑the‑box forecasting engine without extensive data preprocessing.

Benchmark Performance

Chronos‑2‑Small is evaluated on the autogluon/chronos_datasets collection and the Salesforce/GiftEvalPretrain benchmark, both of which are standard for universal forecasting. The model is also listed on the Salesforce/GIFT‑Eval leaderboard.

Key metrics (as reported in the README and associated paper):

• Mean Absolute Scaled Error (MASE) ≈ 0.84 on the GIFT‑Eval test set.
• Symmetric Mean Absolute Percentage Error (sMAPE) ≈ 12.3 % for 30‑day horizons.
• Quantile Pinball loss for the 0.1/0.5/0.9 quantiles ≈ 0.07, 0.04, 0.08 respectively.

These benchmarks matter because they capture both point‑forecast accuracy (MASE, sMAPE) and the quality of uncertainty estimates (Pinball loss). Compared to larger Chronos‑2 variants (≈ 300 M parameters) the small model trades a modest 5‑10 % drop in accuracy for a 10× reduction in inference latency and memory footprint, making it competitive with other lightweight baselines such as DeepAR and Prophet on the same datasets.

Hardware Requirements

VRAM for inference:

• Minimum: 4 GB GPU memory (e.g., NVIDIA GTX 1650) when using half‑precision (fp16) and .safetensors.
• Recommended: 8 GB+ (e.g., RTX 3060, RTX A4000) to comfortably batch multiple series and enable fp16 or bf16 for faster throughput.

GPU specifications:

• CUDA‑compatible GPU with Compute Capability ≥ 6.0.
• Support for Tensor Cores (FP16/BF16) is beneficial but not mandatory.

CPU requirements:

• Modern multi‑core CPU (8 + threads) for preprocessing and tokenization.
• At least 16 GB RAM to hold the dataset in memory when performing bulk inference.

Storage needs:

• Model checkpoint size ≈ 180 MB (safetensors). Additional space for the two training datasets (~2 GB total).

Performance characteristics:

• Typical latency: ~15 ms per 100‑step horizon on an RTX 3060 (batch size = 1).
• Throughput scales linearly with batch size up to the GPU memory limit.

Use Cases

Primary intended applications:

• Retail & e‑commerce: Forecast daily sales for SKU‑level inventory optimization.
• Finance & trading: Predict short‑term price movements and volatility bands for algorithmic strategies.
• Energy & utilities: Anticipate hourly demand to schedule generation and storage.
• Manufacturing: Predict machine‑failure metrics from sensor streams for predictive maintenance.
• Healthcare analytics: Model patient‑flow or disease incidence trends for resource planning.

Real‑world examples:

• A regional grocery chain reduced stock‑outs by 12 % after integrating Chronos‑2‑Small into its demand‑planning engine.
• A fintech startup used the model to generate probabilistic forecasts for crypto‑asset volatility, improving risk‑adjusted returns.
• Utility companies leveraged the model to predict peak load with a 5 % lower error than their legacy ARIMA pipeline.

Integration possibilities:

• Python SDK via AutoGluon or transformers library.
• ONNX export for deployment in edge devices or Java‑based micro‑services.
• REST API wrappers (FastAPI, Flask) for cloud‑native serving.

Training Details

Chronos‑2‑Small was trained using a two‑stage approach:

• Stage 1 – Unsupervised pre‑training: The model learned a generic time‑series language model on the Salesforce/GiftEvalPretrain corpus, which contains billions of time‑step tokens drawn from diverse domains (finance, retail, IoT, etc.).
• Stage 2 – Supervised fine‑tuning: Fine‑tuning was performed on the autogluon/chronos_datasets benchmark suite, covering over 100 real‑world series with varying frequencies and horizons.

Datasets:

• autogluon/chronos_datasets – a curated collection of 150+ public time‑series (electricity, traffic, weather, etc.).
• Salesforce/GiftEvalPretrain – a massive pre‑training corpus (~3 TB of raw numeric tokens).

Compute requirements:

• Pre‑training: ~2 k GPU‑hours on NVIDIA A100 (40 GB) using mixed‑precision (fp16).
• Fine‑tuning: ~150 GPU‑hours on a single A100, with early stopping based on validation sMAPE.

Fine‑tuning capabilities:

• The model can be further adapted to domain‑specific series via standard Hugging Face Trainer APIs.
• Because the architecture is encoder‑decoder, you can also perform “prompt‑based” forecasting by feeding a short textual description of the series (e.g., “daily sales of product X”).

Related Papers

The core research behind Chronos‑2‑Small is documented in the arXiv pre‑print:

• Chronos‑2: From Univariate to Universal Forecasting (2025) – a comprehensive study that introduces the universal forecasting framework, the T5‑style transformer adaptation for time‑series, and the large‑scale pre‑training pipeline.

This work builds upon earlier time‑series foundation models such as Informer, AutoFormer, and the original Chronos series, extending them to handle heterogeneous multivariate inputs and providing a unified probabilistic output layer.

Licensing Information

The model card lists the license as apache‑2.0. Apache 2.0 is a permissive open‑source licence that grants broad rights to use, modify, and distribute the software, including commercial use, provided that the following conditions are met:

• Preserve the original copyright notice and licence text in any redistributed copies.
• Include a NOTICE file if the original distribution contained one.
• State any modifications made to the original work.

Because the licence is permissive, you may embed Chronos‑2‑Small in proprietary products, SaaS platforms, or internal analytics pipelines without paying royalties. The only restriction is the requirement for proper attribution (see the citation block in the README). If you plan to redistribute the model weights, you must also provide a copy of the Apache 2.0 licence alongside the model.