Explore TimechoAI: The New Timer Time‑Series Large Model Cloud Service Now Open for Beta

Timer series of time‑series foundation models

Timer 1.0 (2023) – data‑governed pure‑decoder architecture, demonstrated few‑shot generalization and multi‑task adaptation for time‑series data.

Timer 2.0 – introduced a two‑dimensional attention mechanism that jointly attends over the time axis and variable axis, enabling longer historical context to improve prediction accuracy.

Timer 3.0 – adopted a generative modeling paradigm that can sample multiple plausible future trajectories for a single input, addressing industrial uncertainty. Pre‑trained on trillions of time points; inference speed ≈20× faster than the comparable Chronos model; achieved >5 million monthly downloads on Hugging Face.

Timer‑3.5 (released March 2026) – scaled to 8.3 billion parameters with an 11.5 k time‑point context window. On the GIFT‑Eval benchmark it set a new state‑of‑the‑art, reducing MASE by 7.6 % and CRPS by 13.2 % relative to prior versions, becoming the first billion‑parameter time‑series foundation model.

Timer was developed by the THUML team at Tsinghua University and received the 2025 China Electronics Society Natural Science Award (First Class). It has been deployed in enterprises across energy, aerospace, steel, transportation, and smart‑factory domains.

TimechoAI – cloud service for time‑series large models

The platform provides a unified inference service for Timer models and baseline alternatives:

Timer‑3.5 (default, SOTA)

Timer‑3.0 (stable classic version)

Chronos‑2 (Amazon open‑source pre‑trained model)

AutoARIMA (statistical baseline for simple series)

Holt‑Winters (seasonal decomposition method)

Data can be supplied through three supported formats:

Interactive drawing of a curve on a canvas (quick demo).

Manual entry of time‑series points for fine‑grained control.

File upload of CSV or TsFile (the native format of Apache IoTDB/TimechoDB), enabling direct ingestion of historical datasets.

Covariates such as temperature, holidays, or promotional events can be attached to the input series to reflect external influences during prediction.

Built‑in example datasets (air‑quality, ETTh/ETTm transformer data, exchange rates, climate, disease incidence, weather) allow immediate execution of a forecast and visual comparison with ground truth.

Developers can integrate the service via RESTful APIs or the Python SDK, with documentation covering quick‑start and reference usage.

Typical industrial scenarios

Predictive maintenance – early vibration anomaly detection and remaining‑useful‑life estimation for critical equipment.

Industrial production forecasting – output, yield, and process‑parameter prediction to optimize scheduling and reduce inventory.

Energy load forecasting – high‑precision power‑load prediction for grid dispatch, renewable generation variability, and storage operation.

IoT monitoring and anomaly detection – multivariate sensor modeling, automatic abnormal‑pattern flagging, and alert triggering.

Financial time‑series analysis – exchange‑rate, transaction‑volume, and market‑indicator forecasting with anomaly identification.

Benchmark and performance evidence

On the GIFT‑Eval benchmark, Timer‑3.5 achieved overall state‑of‑the‑art performance, with MASE reduced by 7.6 % and CRPS reduced by 13.2 % compared to previous generations. The model’s inference speed is reported to be 20× faster than Chronos under comparable settings.

Code example

本文
约4000字
，建议阅读
5
分钟
基于时序大模型 Timer，面向时序预测与智能分析的 TimechoAI 时序大模型云服务期待您参与内测！