Envision's Tianji Model Ranks First in ECMWF's AI Weather Quest

What the AI Weather Quest measures

According to Envision, the AI Weather Quest evaluates models through continuous, real-time forecasting and verification, which tests how well they perform under varying weather conditions. Beyond the 15-day medium-range horizon, the competition focuses on probabilistic forecasts for days 19 to 25.

Teams predict three variables: weekly mean temperature, mean sea level pressure, and weekly accumulated precipitation. Temperature and precipitation are evaluated over predominantly land-covered grid cells, while mean sea level pressure is assessed globally. The emphasis is on anticipating weekly conditions, and their uncertainty, several weeks in advance.

Teams submit forecasts weekly. Organizers then compare them with the actual weather that follows and publicly update the leaderboard.

Tianji's result and earlier wins

Envision says Tianji competed under the codename Arcana, meaning "mysteries" or "knowledge", and topped the team ranking. It placed first in the variable-averaged score as well as in near-surface temperature, mean sea level pressure and precipitation for the 19 to 25 day forecast window.

The company points to a run of earlier results for the model:

• A top-tier placement in the global WeatherBench 2 evaluation, on par with Google WeatherNext2 Model.
• First place overall in the deterministic forecast track of the AI Weather Forecast Model Demonstration Program organized by China's National Meteorological Center.
• First place in State Grid's AI load forecasting competition and in China Southern Power Grid's business competition, which Envision cites as proof of value in energy applications.

With the latest result, Envision says Tianji has again shown what it calls full-chain capability, from short-term forecasting to subseasonal prediction.

Why subseasonal forecasting is hard

Robert Zheng, head of Envision's Tianji Weather Foundation Model, described the problem in the company's statement. "Subseasonal forecasting, two weeks to two months out, is one of the hardest challenges in meteorology," he said. He added that the longer the horizon, the greater the uncertainty in atmospheric changes, and the higher the demand on a model's ability to identify evolving weather trends.

Zheng said Envision has built on years of practical experience in new energy to develop what it calls the world's first physics-AI weather model. According to him, the system uses petabyte-scale high spatiotemporal resolution data and a "space, air, ground" three-dimensional data network to deliver forecasts across full time scales.

What it means for users

Envision says Tianji runs on high-performance computing and delivers rapid short-term and global forecasts, with significantly improved accuracy for wind speed and irradiance in key regions. The model covers time scales from minute-level nowcasting to subseasonal trend assessment, which the company describes as end-to-end meteorological intelligence for real-time monitoring and long-term planning.

The emphasis on wind speed, irradiance and grid competitions points to the energy sector as a main target for the technology. The results cited come from competitions and from the company's own description, and the source does not give detailed scores or details on commercial availability.