Google Research releases TimesFM-3 multivariate forecasting model with open weights
Google Research released TimesFM-3, a 330-million-parameter open-weight model that forecasts multivariate time series using historical signals plus known future inputs such as promotions and weather. Rather than autoregressive step-by-step decoding, it fills entire future horizons in one pass, which reporting attributes to lower inference latency. The Decoder says it leads pretrained models on Gift-Eval, FEV-Bench, and Time for point accuracy and calibration, with weights on GitHub and Hugging Face. Around the same reporting window, Google also promoted WeatherNext 3, an hourly energy-focused weather model forecasting turbine-height wind, cloud cover, and surface sunlight for traders and grid operators. Available excerpts stop mid-sentence, so how much direct forecasting sacrifices sequential dependency modeling on nonlinear series is not fully documented in these reports.
Google Research releases TimesFM-3 multivariate forecasting model with open weights
Google Research has released TimesFM-3, a 330-million-parameter model that forecasts time series using related signals and known future events such as promotions and weather. The Decoder reports it ranks first among pretrained models on Gift-Eval, FEV-Bench, and Time for point accuracy and calibration, with weights on GitHub and Hugging Face.
Key takeaway
Direct multi-horizon forecasting with open weights is now benchmark-leading while cutting inference cost for operational forecasting pipelines.
What happened
Google Research released TimesFM-3, a 330-million-parameter forecasting model that analyzes time series alongside related data and known future events like sales promotions or weather forecasts, according to The Decoder.
The Decoder reports TimesFM-3 ranks first among pretrained models on Gift-Eval, FEV-Bench, and Time for point accuracy and calibration, with weights on GitHub and Hugging Face. AI News separately reports Google released WeatherNext 3, an hourly model forecasting wind speed at 100 metres, cloud cover, and sunlight for energy traders and grid operators.
Evidence
TimesFM-3 is a 330-million-parameter model that forecasts using related signals and known future events.
The Decoder · attributed
Google Research has released TimesFM-3, a 330-million-parameter model that forecasts time series using related signals and known future events such as promotions and weather.
TimesFM-3 predicts all future time points in a single pass rather than step by step.
The Decoder · attributed
Instead of predicting the future step by step, the 330-million-parameter model fills in all future time points in a single pass, which cuts compute time and reduces
The Decoder reports TimesFM-3 ranks first on Gift-Eval, FEV-Bench, and Time for accuracy and calibration.
The Decoder · attributed
The Decoder reports it ranks first among pretrained models on Gift-Eval, FEV-Bench, and Time for point accuracy and calibration, with weights on GitHub and Hugging Face.
WeatherNext 3 forecasts wind at turbine height, cloud cover, and sunlight with hourly updates.
AI News · attributed
Google's newest AI weather forecasting model predicts wind speed at 100 metres above the ground, roughly the height of a modern wind turbine. It also forecasts cloud cover and how much sunlight reaches the surface, and it updates every hour.
WeatherNext 3 targets energy traders, grid operators, and renewable developers who pay for similar data.
AI News · attributed
Energy traders, grid operators and wind and solar developers already pay other companies for that
Why it matters
Teams in retail, logistics, and energy can evaluate Google's newest forecasting releases against paid vendor data without waiting on closed APIs, but should test nonlinear dependency coverage on their own workloads.
Limits and uncertainties
The Decoder excerpt cuts off mid-sentence, obscuring the full scope of accuracy trade-offs versus autoregressive baselines.
Reporting does not clarify how TimesFM-3 handles complex nonlinear dependencies that sequential models may capture better.
AI News coverage lacks details on WeatherNext 3 architecture, ground-truth validation, and whether it is open-weight or API-only.
Practical implications
Builders can download TimesFM-3 weights from GitHub and Hugging Face to pilot multivariate forecasting with promotion and weather covariates.
Energy operators and renewable developers can compare WeatherNext 3 hourly wind and solar forecasts against existing paid third-party weather data services.
What to watch
Published benchmark comparisons showing how direct multi-horizon forecasting performs against autoregressive baselines on long horizons and nonlinear series.
Disclosure of WeatherNext 3 validation against ground-truth sensor data and its commercial access model for grid operators.
