Envision's Tianji tops ECMWF AI weather contest as China pushes satellite models

A contest for weeks three and four

The AI Weather Quest is not a one-off bake-off. ECMWF runs it as continuous forecasting and verification. Teams submit entries every week, and a public leaderboard updates as results come in. Scoring runs past the 15-day medium-range horizon. The contest covers probabilistic weekly mean temperature and mean sea level pressure, plus weekly accumulated precipitation, for days 19 to 25. Temperature and precipitation are scored over mostly land-covered grid cells. Pressure is scored globally. The design rewards a model for describing a week three weeks out.

That matters to anyone buying power, not just meteorologists. Subseasonal outlooks for weeks three and four feed maintenance, hedging and trading decisions at utilities, wind and solar operators and insurers.

What is unknown is independent skill at site level, and what access to Tianji costs. Envision says Tianji took first place across all assessed metrics, including variable-averaged, near-surface temperature, pressure and precipitation. The claim appears in a company release dated 30 October. Yonhap carried it with a note that it was not edited.

China's satellite models target the short end

China used the same week to show a different approach. It unveiled two AI models for extreme weather at the 2026 Fengyun Satellite User Conference in Changsha, Hunan province, on Thursday.

The first interprets typhoon cloud images. According to China Daily, it integrates more than 20 types of real-time satellite products, analyses images every 15 minutes and produces a text report within 27 seconds. That report covers a storm's location, intensity, structure and likely evolution. Han Bowei, an engineer at the National Satellite Meteorological Center, said a typhoon's three-dimensional thermal structure is one of the factors behind intensity change. The second model handles severe convective weather, the fast-growing thunderstorm systems that bring damaging winds, hail and flash floods. Han said it uses two hours of continuous Fengyun 4 data and can predict occurrence, development and intensity four hours ahead. A study on it appeared last year in PNAS.

Reach is the point. The China Meteorological Administration says China has 10 Fengyun satellites in orbit and shares their data with 141 countries and territories, Sharjah24 reported. Han said the two models are the first in a planned series covering typhoons, rainstorms, sandstorms, heavy fog and cold waves.

Why the 19 to 25 day window is hard

Two weeks to two months out is where forecasting gets uncomfortable. Robert Zheng, who heads Envision's Tianji Weather Foundation Model, called subseasonal forecasting over that range one of the hardest challenges in meteorology.

Longer horizons raise uncertainty, he said. They demand more from a model's ability to identify evolving weather trends. Envision calls Tianji the world's first physics-AI weather model, drawing on petabyte-scale data and a three-dimensional network spanning space, air and ground. It says Tianji placed top-tier in WeatherBench 2, level with Google's WeatherNext2. It also won the deterministic track of a demonstration programme run by China's National Meteorological Center. Grid work is the commercial hook. Envision says Tianji won first place in State Grid's AI load forecasting competition and in China Southern Power Grid's business competition. It also says the model improves accuracy for wind speed and irradiance in key regions. Those results come from the company.

Open models take a different route

Nvidia is selling the opposite model: open weights and tools. Its Earth-2 family lets developers take models and customise them with their own data.

The newest addition is StormScope EU. It is a regional version trained on European weather and climate data. It uses EUMETSAT observations to predict evolving cloud conditions across 16 Meteosat channels at 10-minute intervals. Nvidia showed a live pipeline at GTC Berlin with Earthmover. AXA is using FourCastNet 3 to study regional cyclone risk beyond the historical record, across more than 1,000 simulated seasons. Vaisala Xweather is combining Earth-2 with its own observations and lightning data. TotalEnergies is evaluating StormScope-based solar-radiation nowcasts at two photovoltaic sites.

The University of Manchester took a different route. Researchers there adapted the CorrDiff model for UK air-quality work. They trained it on the Isambard-AI supercomputer and then ran it locally on Nvidia DGX Spark systems.

What is still unproven

The scoreboard is public, but the 30 October announcement was Envision's. No independent evaluation of Tianji's skill at a wind farm or in a control room has been published.

That gap matters because the contest measures weekly means over grid cells. A grid operator needs hourly wind and irradiance at a specific site. A bad ramp costs money whether or not the mean was right. Envision says Tianji spans minute-level nowcasting to subseasonal trends, but the Quest result speaks only to the far end. Cost is the other blank: Envision has not said what access to Tianji costs. On the satellite side, Han said all conclusions from the Fengyun models rest on satellite observations and can be verified. That is a claim about traceability rather than accuracy in a live warning chain.