WeatherNext 2 Open Source: A Cyclone Model Teens Can Run

WeatherNext 2 open source is the part of this story that matters for a curious teenager. In August 2026, Google released the model to the research community, having already published in Nature that it predicts “cyclone track, intensity, and wind structure with state-of-the-art accuracy,” which Google characterized as “a decade of meteorological progress in one model.” The technical details are learnable by a motivated 15-year-old: it produces hundreds of possible weather futures from one starting snapshot, each one in under a minute on a single TPU chip, out to 15 days, at one-hour granularity, and it outperforms the previous WeatherNext on 99.9% of variables.

Key Takeaways

• Architecture: a Functional Generative Network (FGN), which injects noise directly into the model’s internals rather than adding it to the output, so the ensemble members stay physically coherent.
• Speed: 8x faster than its predecessor. Each forecast takes under a minute on a single TPU. Traditional physics ensembles need supercomputers and hours.
• Range: 0 to 15 days, at 1-hour granularity, with hundreds of scenarios generated from a single input.
• Accuracy: surpasses the previous WeatherNext on 99.9% of variables; cyclone track, intensity, and wind structure results were published in Nature.
• Access: open-sourced to the research community (August 2026), plus Earth Engine, BigQuery, and Vertex AI early access. It also powers forecasts in Search, Gemini, Pixel Weather, and Google Maps.

What an ensemble forecast is, and why one forecast is a lie

A single weather forecast is a point estimate of something fundamentally uncertain. The atmosphere is chaotic: microscopic differences in today’s measurements produce large differences a week out. Lorenz established this in the 1960s and it hasn’t changed.

The professional response is ensembles. Run the forecast many times from slightly different starting conditions and slightly different model settings. If 80 of 100 runs put the hurricane over Florida, that’s an 80% probability, and it’s a better answer than any single run. This is where “40% chance of rain” comes from.

The cost is brutal. A traditional physics-based ensemble means solving fluid dynamics equations across a global grid, dozens of times, on a supercomputer, for hours. Which limits how many members you can afford and how often you can rerun.

That constraint is what WeatherNext 2 attacks.

How the WeatherNext 2 open source model actually computes a forecast

What the model learned from. Decades of reanalysis data: the global record of atmospheric state (temperature, pressure, wind, humidity) on a grid over time. Training means learning to map the atmosphere’s state now to its state later, from millions of such pairs.

The problem with naive ensembles in a neural model. The obvious way to make a neural network produce multiple scenarios is to perturb the input or add randomness to the output. Both break things. Perturbing inputs is expensive and the variation doesn’t reflect real forecast uncertainty. Adding noise to outputs produces forecasts where the temperature field and the wind field disagree with each other, because the noise was applied independently to each variable. You get a forecast that is physically impossible.

What the FGN does instead. It injects noise inside the model architecture. Google’s description: this approach means “the forecasts it generate remain physically realistic and interconnected.” Because the randomness enters early and propagates through the whole network, every variable in a given ensemble member is shaped by the same perturbation. The temperature and the wind stay consistent with each other.

Why that phrasing about marginals and joints matters. Google says the system learns individual weather elements (marginals) during training and then forecasts complex interdependent systems (joints). Getting the marginal right means “the temperature in Miami is probably 82°F.” Getting the joint right means “if the temperature in Miami is 82°F, then the wind in Tampa is probably from the southwest and the pressure gradient is this.” Cyclones are joint phenomena. You cannot forecast a hurricane by forecasting each grid point independently.

What it got right. 99.9% of variables improved over the previous model. Cyclone track, intensity, and wind structure at state-of-the-art accuracy, published in Nature. Hundreds of ensemble members in under a minute each on one TPU, 8x faster than before.

What it can’t do. It’s a 15-day model, so seasonal and climate-scale questions are out of scope. It learned from historical reanalysis, so an atmosphere that drifts outside that historical distribution is a genuine open risk for any learned weather model. And it forecasts the atmosphere, not impacts: whether a cyclone floods your town depends on drainage, terrain, and tide, which are separate models entirely.

The analogy: rolling one weighted die versus many

Imagine predicting tomorrow’s lunch line length. You could give one number: 14 people. Or you could give a distribution: usually 10 to 18, occasionally 25 on pizza day.

Now the hard part. Suppose you also want to predict how long the line is at the second cafeteria, and the two are related — if line A is long, line B is short, because people walked over. A bad ensemble generates line A and line B independently, and sometimes produces a world where both lines are empty and nobody ate lunch. A good ensemble generates scenarios, each internally consistent.

The FGN’s trick is to put the randomness at the start of the story, so each scenario unfolds coherently, rather than sprinkling randomness on the final numbers.

AI weather models, compared

Model	What it does	Speed	Ensemble	Availability
Traditional physics (ECMWF IFS, NOAA GFS)	Solves fluid dynamics on a global grid	Hours on a supercomputer	Dozens of members	Public forecasts; NOAA data is free
GraphCast (2023)	Learned graph neural network, deterministic	Under a minute	Single forecast	Open weights
WeatherNext 2 (2026)	FGN, probabilistic ensembles, 0–15 days, 1-hour steps	Under a minute per member on one TPU	Hundreds of members	Open-sourced to researchers; Earth Engine, BigQuery, Vertex AI
WeatherNext 3	Successor with 5 km resolution for temperature and humidity, 10 km for wind, hourly forecasts	Not published	Not published	BigQuery, Earth Engine, Google Maps Platform, Cloud Storage
Hydrology models (separate)	What happens after rain hits the ground	Varies	Varies	NOAA National Water Model

The bottom row matters for anyone thinking about floods. A great atmospheric forecast still needs a water model to say what happens on the ground, which is covered in our piece on the hybrid hydrology model at Penn State.

How to Teach Your Kid About Ensemble Weather Forecasting

Ages 5–8: The Guess Jar With Many Guesses

Ask your kid to predict tomorrow’s high temperature. Write it down. Then ask four other family members. Now you have an ensemble. Tomorrow, check who was closest and, more importantly, note whether the real answer fell inside the range of all the guesses. That’s exactly how forecasters judge an ensemble: not on any single member, but on whether the truth landed inside the spread.

Ages 9–12: Build a Probability Forecast

For two weeks, have your kid record the forecast probability of rain each morning and whether it actually rained. At the end, group the days: on the days forecast at 30%, did it rain about 3 times out of 10? That’s called a reliability check, and it’s how professional forecasts are scored. Most kids are surprised that a 30% forecast is supposed to be wrong most of the time.

Ages 13+: Open Real Forecast Data

This is where the open-source release becomes usable. WeatherNext data is accessible through Google Earth Engine and BigQuery, both of which have free tiers for research and education, and NOAA publishes its own model output freely. Have your teen pull a multi-day forecast for one location, plot the ensemble spread, and compare it to what actually happened. Then have them find a case where the ensemble was wide (high uncertainty) and one where it was narrow, and explain the weather situation behind each.

The question to ask: “If the model gives you 200 different forecasts, which one is the real prediction?”

What to actually do at home

Teach spread, not just the number

The most valuable weather literacy skill is reading uncertainty. “Highs near 78, but the models disagree by 10 degrees” is a different day than “highs near 78, all models agree.” Kids who learn to ask “how confident?” become adults who read every forecast better, including the non-weather ones.

Use hurricane season as the teaching window

Cone-of-uncertainty graphics on the news are ensemble spread, visualized. Pull up the National Hurricane Center during an active storm and explain that the cone is not the size of the storm; it’s the range of possible centers. That single correction fixes a very common misunderstanding.

Connect speed to access

Under a minute on one TPU instead of hours on a supercomputer is what makes it plausible for a university lab, a national weather service in a small country, or an advanced high school student to run experiments. Cost per forecast is a democratization story, and it’s worth naming explicitly.

Pair it with a physical instrument

Recording your own temperature and rainfall makes forecast verification concrete. Our rain gauge and weather station build gives you a local measurement to check forecasts against, which is more engaging than checking someone else’s number.

What not to do

Don’t tell your kid the AI “understands” weather. It learned a mapping from atmospheric states to later atmospheric states, with a clever noise-injection design that keeps variables consistent. It contains no equations for fluid dynamics. That’s both the source of its speed and the reason researchers still keep physics models running alongside it.

What to Watch For Over the Next 3 Months

• Week 4: Your kid can explain what a 30% chance of rain means and why being wrong 7 times out of 10 is correct behavior for that forecast.
• Month 2 red flags: They treat a 10-day forecast as reliable. Have them log day-10 forecasts against reality for two weeks; the error will teach the lesson better than you can.
• Month 3 self-check: During any storm, ask them to find the ensemble spread rather than the single track. If they go looking for the cone instead of the line, the concept landed.

Frequently Asked Questions

What does “open source” mean for WeatherNext 2?

Google released the cyclone model to the research community in August 2026 to support global climate resilience work. In practice, access to WeatherNext outputs also runs through Earth Engine, BigQuery, and Vertex AI early access. Open release means outside researchers can examine and build on the model rather than only consuming its forecasts.

Can a teenager actually run this?

Running the model’s training is out of reach; it needs serious hardware. But working with its output is genuinely accessible. Earth Engine and BigQuery have free tiers, NOAA publishes model data openly, and plotting an ensemble spread from real forecast data is a weekend project for a teen comfortable with a spreadsheet or basic Python.

How is this different from GraphCast?

GraphCast, from 2023, produces a single deterministic forecast using a graph neural network. WeatherNext 2 produces probabilistic ensembles: hundreds of internally consistent scenarios, using the Functional Generative Network architecture, at 1-hour granularity out to 15 days. The shift from one forecast to a distribution of forecasts is the substantive change.

Do meteorologists still need physics models?

Yes. Learned models are trained on the historical atmosphere, so their behavior in genuinely unprecedented conditions is not guaranteed. Physics models generalize because they encode conservation laws. Most operational centers run both and compare, which is the sensible engineering answer.

Does this mean hurricane forecasts will get better?

The Nature results on cyclone track, intensity, and wind structure are real improvements in forecast skill. Whether that reaches your family depends on national weather services adopting these models operationally and on emergency managers changing procedures, both of which move slower than the research.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Google DeepMind. “WeatherNext 2: our most advanced weather forecasting model.” https://blog.google/technology/google-deepmind/weathernext-2/
2. Google. (2026, August). “Google AI updates, August 2026” (WeatherNext 2 open-sourcing; Operation Blue Skies; Gemma downloads). https://blog.google/innovation-and-ai/technology/google-ai-updates-august-2026/
3. Google DeepMind. WeatherNext model family and access methods. https://deepmind.google/models/weathernext/
4. NOAA National Weather Service. Forecast products and model data. https://www.weather.gov/
5. NOAA National Hurricane Center. Track forecast cone explanation. https://www.nhc.noaa.gov/
6. NOAA Office of Water Prediction. “About the National Water Model.” https://water.noaa.gov/about/nwm
7. Google Earth Engine. Data catalog and free research access. https://earthengine.google.com/