Envision’s Tianji tops ECMWF’s AI Weather Quest: a physics-AI model aimed at the 19–25 day window

Envision announced on October 30, 2026 that its weather foundation model, Envision Tianji, ranked first overall in the latest edition of the AI Weather Quest, the artificial-intelligence weather competition organized by the European Centre for Medium-Range Weather Forecasts (ECMWF). Tianji outperformed nearly 20 top teams worldwide, taking first place across all assessed metrics for the 19–25 day forecast window: variable-averaged near-surface temperature, mean sea level pressure, and precipitation. The model topped the team ranking under the codename Arcana, which the announcement says means “mysteries” or “knowledge.”

The hard window is weeks three and four

The AI Weather Quest is built to test staying power, not a single good forecast. Teams submit forecasts weekly, and organizers evaluate them against the weather that actually follows, updating a public leaderboard as results come in. The competition looks beyond the 15-day medium-range horizon and scores probabilistic forecasts: weekly mean temperature and mean sea level pressure, and weekly accumulated precipitation, for days 19–25. Temperature and precipitation are assessed over predominantly land-covered grid cells, while mean sea level pressure is assessed globally — a design that rewards models that can anticipate weekly conditions, and their uncertainty, several weeks in advance.

“Subseasonal forecasting — two weeks to two months out — is one of the hardest challenges in meteorology,” said Robert Zheng, head of Envision’s Tianji Weather Foundation Model, in the announcement. “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 described Tianji as the world’s first physics-AI weather model, built, in his account, on Envision’s years of work in new energy, petabyte-scale high spatiotemporal resolution data, and a “space–air–ground” three-dimensional data network. That “first” characterization is Envision’s; the announcement offers no independent comparison to settle it.

Why energy cares first

For Envision, a green-technology company, the subseasonal window is operational, not academic. The company says Tianji, powered by high-performance computing, delivers rapid short-term and global forecasts with significantly improved accuracy for wind speed and irradiance in key regions, covering the full range from minute-level nowcasting to subseasonal trend assessment. That span is what grid and renewables planning runs on: knowing, weeks out, whether wind and sunshine will show up, and by how much, shapes maintenance scheduling, power trading, and how much backup capacity to hold.

The AI Weather Quest result is the latest in a run Envision cites for the model. Tianji earned what the company calls a top-tier placement in the global WeatherBench 2 evaluation, on par with Google’s WeatherNext2 model, and first place overall in the deterministic forecast track of the AI Weather Forecast Model Demonstration Program organized by China’s National Meteorological Center. It also won first place in State Grid’s AI load forecasting competition and in China Southern Power Grid’s business competition — results that tie the same forecasting capability to the demand side of the grid, predicting how much power people will use, not just what the sky will do.

What this result does not settle

A competition win is not a product launch, and this announcement should not be read as one. Every score in it is reported by Envision from the competition; the release publishes no error margins, no per-metric breakdown beyond first place across all assessed metrics, and no details on model size, training compute, or how the physics guidance is implemented. Most importantly, the announcement says nothing about whether Tianji is publicly available — no open weights, no API, no service offering is described. Anyone waiting to run the model themselves has, on this announcement alone, nothing to download.

What would strengthen the claim is straightforward: the public leaderboard the competition maintains, independent checks of the same forecast period, and a fuller technical account of the model behind the codename Arcana. Until then, the fair reading is narrow and still notable: in a live, weeks-ahead forecasting contest run by Europe’s weather centre, a model built by an energy company finished first on every metric it was scored on. For a business whose core problem is guessing wind and sun weeks in advance, that is the result that matters — and the next announcement that counts will be the one that says how the rest of the world gets to use it.

Sources: the Envision announcement via PR Newswire, republished by TNGlobal. The editorial illustrations are AI-generated.