Envision says its weather AI took first place in ECMWF’s three-week forecast contest

Envision says its Tianji weather AI has won a test of forecasting roughly three weeks ahead. In an October 30, 2026 announcement, the company claimed first place overall in ECMWF’s AI Weather Quest. The result concerns weekly weather conditions and their uncertainty—not a promise of precise daily forecasts that far into the future.

Competing under the codename Arcana, Tianji outperformed nearly 20 teams, according to Envision. The company said it led every assessed metric for the 19–25-day forecast window: the average across weather variables, near-surface temperature, mean sea level pressure and precipitation. Those are several measures of performance within one defined forecasting task.

A week of conditions, three weeks away

The AI Weather Quest is organized by the European Centre for Medium-Range Weather Forecasts, or ECMWF. Envision’s description puts the task beyond the 15-day medium-range forecasting horizon. Entrants produce probabilistic forecasts: estimates of possible outcomes and their likelihood, rather than one fixed prediction. The target is a week’s overall conditions during days 19 through 25.

• Temperature: the weekly mean near the surface, evaluated over grid cells that are predominantly land.
• Precipitation: the accumulated total over the week, also evaluated over predominantly land-covered grid cells.
• Mean sea level pressure: the weekly mean, assessed globally rather than through the land-focused coverage used for temperature and precipitation.

Teams submit forecasts each week. Organizers then compare those predictions with the weather that actually occurs and publicly update the leaderboard. This rolling process assesses sustained performance under changing weather conditions. The forecasts are submitted before the observations used to judge them, making the timing of the test an important part of its design.

Robert Zheng, head of Envision’s Tianji Weather Foundation Model, describes subseasonal forecasting as covering two weeks to two months ahead. He calls it one of meteorology’s hardest challenges. His explanation focuses on the increasing uncertainty that a model must handle as it looks further forward.

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.

Robert Zheng, head of Envision’s Tianji Weather Foundation Model

The model behind the claim

Envision describes Tianji as a physics-AI weather model. Zheng says it draws on petabyte-scale data with high spatial and time resolution, alongside a three-dimensional data network spanning space, air and ground. The company positions that combination as a way to identify evolving weather patterns across multiple forecasting horizons, rather than serving only the competition’s three-week window.

Its claimed scope runs from minute-level forecasts to subseasonal trend assessment. Envision also says high-performance computing enables rapid short-term and global forecasts, with improved accuracy for wind speed and solar irradiance in key regions. Those broader capabilities are company claims, distinct from the specific AI Weather Quest result.

The announcement also cites earlier weather-model evaluations. Envision says Tianji achieved a top-tier WeatherBench 2 placement, on par with Google’s WeatherNext2 model, and ranked first in the deterministic forecasting track of a program organized by China’s National Meteorological Center. These provide performance history, not separate confirmation of the latest ranking.

The ranking remains an issuer-reported result. A November 2 republication on Japan Wire by Kyodo News carries the same Envision release and explicitly identifies it as issuer-provided material; it is not independent corroboration. Envision’s announcement gives the claimed placements but no numerical scores, winning margins or evaluation duration. Those missing details leave the size and persistence of Tianji’s advantage unresolved.