AINEWS
The release produces a spread of plausible forecasts instead of a single answer, matching how traditional weather agencies work and narrowing a credibility question ML weather has long struggled with.
Aurora 1.5, the open weather foundation model Microsoft Research first released in 2024, now produces a spread of plausible forecasts rather than a single answer, the change Microsoft itself frames as the model's most consequential update. The shift puts the open model on the same footing as traditional weather agencies that have issued probabilistic outlooks for decades, and it directly answers a long-standing criticism of machine-learning weather tools: that they look confident even when the atmosphere isn't.
The release adds three things on paper. The model now outputs 22 new weather variables on top of the original Aurora foundation, covering indicators relevant to energy, agriculture, transport, and climate risk. It produces forecasts at hourly cadence rather than the coarser rhythm of the original release. And it ships ensemble forecasting capability, described by Microsoft as motivated by "forecast sensitivity to initial conditions and model uncertainty." Two of those are vendor-spec dressing. The third is the mechanism that changes who can trust the answer.
Ensemble forecasting runs the same model many times with slightly perturbed starting conditions to capture how much the answer wobbles when the inputs shift. A single ML weather run is brittle: one number, no way to know if the model is sure. A spread of runs gives a downstream user a range, and the range is the decision. A 90% chance of a wind lull is not the same answer to a grid operator as a 70% chance. A 60% chance of frost is a different decision for a farmer than a 30% chance. The U.S. National Weather Service and the European Centre for Medium-Range Weather Forecasts have issued probabilistic outlooks in this style for years. The capability had been a missing piece for the open Aurora line, and the new release closes that gap.
The original Aurora model was published in Nature in 2025 and showed it could be fine-tuned for medium-range weather, ocean waves, atmospheric chemistry, and high-resolution regional forecasting. The model and its checkpoints were released as open source on GitHub and Hugging Face. The credibility question has always been operational: whether the openly available weights actually forecast as well as the systems that drive daily decisions in the sectors Microsoft is now targeting, or whether the production version lives behind a managed-access tier the open release does not include. Aurora 1.5 narrows the capability gap on paper. It does not, by itself, close the credibility one.
The release also blurs a line between two tiers that the announcement describes separately. Researchers can download the model and checkpoints from Microsoft's GitHub and Hugging Face repositories and run them locally. The operational tier, Microsoft Weather, is a separate product: data, infrastructure, managed access, and the version a buyer would actually integrate. Microsoft's own framing positions Aurora 1.5 as the bridge from the open research model to the production service, which is honest about the relationship but leaves a reader guessing about which version is doing the work in any given application. The reader has to ask which version a claimed benchmark is using before the number means anything.
Microsoft has framed Aurora 1.5 as a state-of-the-art step for ML weather. Internal benchmark wins on Microsoft's own datasets are exactly what every other ML weather vendor has shown at release, and those wins do not always transfer to operational weather performance once the model meets real-world noise, missing inputs, and the rare-event tails that traditional NWP systems were designed to handle. An independent atmospheric scientist who has benchmarked ML weather models against an established baseline, or a documented case where a probabilistic Aurora 1.5 forecast changed a downstream decision in energy dispatch, irrigation, or flight planning, would settle the question. Neither has appeared yet.
What to watch for in the next release: whether Aurora 1.5's reported improvements hold up under a benchmark from a group that didn't help train it, and whether Microsoft publishes operational results from its own Weather service running the same model. A single number is a research claim. A range, well calibrated, is an operational tool. The release closes one of those gaps and the more important one is still open.