Microsoft Says Its New ML Tool Makes 2-to-6-Week Forecasts Reliable

Microsoft researchers say they've built a machine learning tool that makes weather forecasts accurate two to six weeks out — a range where traditional forecasting has never been dependable. The company claims the approach is also cost-effective, and it backed the claim by publishing the findings today in the journal Nature. The tool itself is already available as open-source software on GitHub.

That combination — a peer-reviewed Nature paper and an immediate open-source release — matters. It's not a product demo or a benchmark tweet. It's a research result other scientists can check, and a piece of code anyone can run.

Why the two-week wall matters

Old-school physics-based weather forecasting and even new AI models typically become less reliable beyond two weeks, Microsoft says. That's the problem the researchers claim to have solved using machine learning.

The stakes are practical, not academic. Farmers and emergency responders are the two groups Microsoft names as beneficiaries, because both need to plan ahead with greater certainty than a ten-day outlook allows. A farmer deciding when to plant, or an agency staging supplies before a storm season, is working on exactly the two-to-six-week horizon this tool targets.

The takeaway: if the claim holds, the useful forecast window roughly triples — from about two weeks to as far as six.

What did Microsoft actually build?

The answer is: not another weather model. According to the company, the researchers didn't create a new standalone weather model. The source material is sparse on the technical mechanism beyond that, but the framing is clear — Microsoft took a different route from the generative-forecasting approach several of its competitors have chosen.

That's a meaningful distinction in the current race. The dominant AI-weather playbook has been to train a large model to emulate atmospheric dynamics directly, replacing chunks of the physics simulation. Microsoft's researchers instead aimed at the sub-seasonal gap — the two-to-six-week range where both physics-based systems and those newer AI models degrade.

The tool is open source now, hosted publicly on GitHub, so the method is inspectable from day one.

The takeaway: Microsoft positioned this as a fix to a known forecasting gap, not as a rival to the big end-to-end weather models.

Who else is racing to rebuild forecasting with AI?

Microsoft is not alone in this market. The company joins Google, Nvidia, and other firms developing AI tools for weather forecasting.

• Google has been working on AI weather forecasting, including satellite-data-driven model updates.
• Nvidia launched AI technologies aimed at weather forecasting in January 2026.
• A wider field of companies, as Fast Company has reported, is changing how weather forecasting gets done with AI tools.

That competitive context explains why Microsoft chose Nature and open source as its launch channels. In a crowded field, verifiability is the differentiator: a claim about six-week forecasts is easy to make and hard to check unless the code and the paper are both public.

The takeaway: AI-driven forecasting has become a multi-company race, with Microsoft, Google, and Nvidia all pushing on different parts of the problem.

What happens next?

For now, the result is a research publication and a public codebase, not a shipped forecasting product. The questions that will decide its real-world impact are the obvious ones: whether independent forecasters adopt the tool, and whether the two-to-six-week accuracy gains hold up outside the researchers' own evaluations.

The open-source release makes that next step possible — any meteorological service or research group can now test the claim directly.