Jua AI

What is Jua AI?

Jua AI is a Swiss company that forecasts global weather with its own fully AI-driven model instead of correcting traditional physics forecasts. Its homepage cites 1 km² detail, five-minute steps and better accuracy at 48 hours than leading numerical models manage at 12 hours.

The engine belongs to a family called the Earth Physics Transformer, or EPT. Jua’s July 2025 technical report describes EPT-2 as producing hourly global forecasts up to 20 days ahead, run four times a day, on a grid of roughly 9 km at the equator. The finer 1 km² figure on the homepage therefore appears to describe the product output rather than that published base model.

Jua was founded in Zurich in 2022 by Andreas Brenner and Marvin Gabler and has reportedly raised $27 million. The company says its team has published more than 100 research papers, and its platform, API and Python SDK were all available in October 2026.

Key Features

With the EPT-2 model family and a hindcast archive, energy teams can test Jua’s AI forecasts against history before trusting them with money.

• EPT-2 forecasts: The flagship model produces global hourly forecasts out to 240 hours from one unified system. Jua’s report says it beats ECMWF’s IFS HRES beyond the 12-hour mark, although those benchmark results were written by Jua itself.
• Ensemble version: EPT-2e generates a set of perturbed forecasts so you can see a likely range rather than one line. The technical report tested 10 members, while a data marketplace lists 30, so ask which configuration your contract includes.
• Rapid refresh: A faster edition is listed as updating every hour, or 24 times a day, instead of every six hours. Shorter refresh cycles matter most for intraday power trading, where a single wind ramp can move prices.
• Hindcasts for backtesting: Historical runs of Jua’s models are available so you can replay past periods and compare results with what actually happened. The SDK exposes hindcasts directly, and the marketplace listing says deeper history is provided on request.
• Third-party models: The API reportedly also serves Microsoft Aurora, Google DeepMind’s GraphCast and conventional numerical forecasts. Having several models behind one key makes side-by-side comparison easier for analysts building blended forecasts.
• Market aggregates: The Python SDK can return weather and power data aggregated to energy market zones across several countries. That saves traders from summing hundreds of grid points themselves before feeding results into pricing models.

Pros

Energy traders pick Jua for fast, energy-focused AI forecasts delivered as data, with published research and backtesting tools to check the claims.

• Fine time steps: Five-minute output and hourly refreshes suit markets that react within the hour. Intraday traders and grid balancing teams gain the most, since they act on changes long before the next standard model run.
• Published research: Jua released a public technical report on arXiv describing EPT-2’s design, training cost and benchmarks. Few commercial forecast vendors share that much, which gives data scientists something concrete to scrutinise.
• Efficient training: The report says EPT-2 trained on 8 H100 GPUs over 10 days, against far larger budgets for comparable models. Lower compute cost may let Jua refresh and retrain models more often than heavier rivals.
• Developer-friendly access: An open-source MIT-licensed Python SDK and a REST API with an interactive test page shorten integration work. Results arrive as xarray datasets that drop straight into pandas or matplotlib.
• One global model: A single model covers the whole planet, so you avoid stitching regional forecasts together. Firms expanding into data-poor regions, mountains or coastlines benefit, as Jua highlights these as target use cases.

Cons

Jua publishes no price figures, so every plan starts with a sales conversation.

• No public prices: You cannot estimate costs without contacting sales, and no free tier is advertised. Smaller firms and independent researchers may find that hard to justify before a trial shows results.
• Self-reported accuracy: Headline claims against ECMWF and Aurora come from Jua’s own report and homepage. Independent verification is limited, so run hindcasts for your own sites before relying on the numbers.
• Inconsistent specifications: Public sources disagree on details such as ensemble size and resolution. The homepage cites 1 km², while the technical report gives about 9 km, so confirm specifications in writing.
• Specialist audience: Jua is not a consumer weather app for checking tomorrow’s rain. Most value appears only when forecasts feed trading models, grid tools or a weather product you already run.
• Technical set-up: The SDK needs Python 3.11 or later and some data engineering to reach full use. Teams without analysts may lean on the web platform, which offers less flexibility than code.

Who is Using Jua AI?

Energy traders and renewable operators are Jua’s core users, with weather services and utilities following because they need frequent, high-resolution forecasts.

• Energy traders: Trading desks price power and gas against expected wind, solar and temperature. They use frequent refreshes to react to forecast changes before the wider market does.
• Renewable operators: Wind and solar owners forecast generation for dispatch and scheduling. Jua’s power-focused outputs help them plan output and reduce imbalance charges.
• Grid and utility teams: Network planners forecast demand and load during heat waves or cold snaps. Ensemble ranges help them prepare for less likely but costly scenarios.
• Weather services: B2B and consumer forecast providers add Jua data to their own products. The homepage names coastal and mountain terrain as areas where it aims to improve on older models.

Pricing

Jua AI publishes no prices, so every plan is quoted after you contact the company (checked October 2026 on the official site and a data marketplace).

Futurepedia and Neura describe pricing as custom and note no free tier, while the InsightCommodity marketplace calls the API pay-as-you-go, without giving figures.

Disclaimer: Weather data contracts depend on regions, variables and query volume, so treat any estimate as provisional until Jua sends a written quote.

What Makes Jua AI Unique?

Jua replaces the standard numerical-model chain with one end-to-end AI model built specifically for energy markets. Most commercial providers still start from government models such as GFS or ECMWF IFS and adjust them with machine learning, whereas Jua says its model is not based on either.

The second difference is openness about method. Jua published a technical report with training costs and benchmarks, and it ships an MIT-licensed SDK. It also offers rival models such as Aurora and GraphCast through the same API, so customers can compare rather than take one vendor’s word for it.

Compatibilities and Integrations

Jua connects through a REST API and a Python SDK, with data returned in formats that analysts already use.

• Platforms: A browser-based forecasting platform for viewing and comparing models without code.
• API: A documented REST API, now on version 2, with an interactive endpoint test page at query.jua.ai.
• SDK: An open-source Python package installed with pip or uv, authenticated through the browser or an API key from the developer dashboard.
• Data formats: JSON, Xarray and Zarr are reported, and SDK results work directly with pandas and matplotlib.
• Documentation: Guides at docs.jua.ai cover forecasts, hindcasts and market aggregates, with Markdown versions of each page.

Jua AI Tutorials

A practical first test takes a few days and focuses on backtesting Jua against the forecast source you already use.

1. Request a trial or demo from Jua and agree the regions, variables and models you need.
2. Create an API key in the developer dashboard, or run the browser sign-in from the SDK.
3. Install the SDK with pip on Python 3.11 or later.
4. Pull an EPT-2 forecast for a site or market zone you know well.
5. Request hindcasts for the same location and compare them with observed data and your current provider.
6. Only then connect the feed to trading or dispatch tools, starting with a small, monitored position.

How We Rated It

Opaque pricing and largely self-reported accuracy keep Jua AI at 3.9, even with strong model performance and developer access.

Summary

Jua AI is a good fit for energy traders, renewable operators and weather businesses that want fast AI forecasts as structured data. Its research, hindcasts and multi-model API give technical teams real tools to check its claims. The main limitations are hidden pricing and specifications that differ between sources. Casual users will find nothing here, since it is built for data pipelines. Before signing, run a hindcast comparison for your own sites over at least one volatile season.