For most of the last decade, AI improved almost everything except the thing most businesses actually need: a reliable prediction of what a number will be next week, next month, or next quarter. Text generation improved dramatically. Image recognition improved dramatically. But forecasting, the core task behind inventory decisions, energy procurement, capacity planning, and financial budgeting, kept being dominated by old statistical formulas that a competent analyst could have written in Excel.
That gap has been closing over the past year, and Datadog's release of Toto 2.0 this week is the clearest evidence yet of how fast.
Datadog, a company primarily known for software monitoring, has published a family of five AI forecasting models, ranging from a small version suited for edge deployments to a 2.5 billion parameter flagship. All five are free to use, modify, and commercialise under an open Apache 2.0 license. The weights, meaning the fully trained model files, are available on Hugging Face right now.
The single most important result from Toto 2.0 is not a benchmark score. It is the proof that making the model bigger reliably makes it better, with no ceiling visible yet. Every prior attempt to scale this class of model hit a wall. Toto 2.0 is the first family where size consistently translates to accuracy, in the same way that larger language models reliably write better than smaller ones. That was not true of time series AI a year ago.
What makes this particularly striking is the training data. Toto 2.0 was trained almost entirely on technology monitoring data, the kind of readings that come from servers, networks, and software systems. It saw essentially no public retail, energy, financial, or weather forecasting data before testing. Yet it outperformed every competitor on a 97-dataset general-purpose benchmark spanning all of those domains. A model that learned from one type of signal generalising to completely different signals is the kind of cross-domain transfer that, until recently, only language models had shown.
For a procurement manager, a demand planner, or an operations director, the relevant question is: what does this mean in practice? The honest answer is that this technology is still one or two steps from plug-and-play for most businesses. The models are open and free, but they require a technology team to connect them to your data and integrate them into a workflow. That is not a trivial task.
However, the competitive clock is ticking. Companies like IBM, Amazon, Salesforce, and now Datadog are all racing to build this capability, and the enterprise software vendors that serve industries like retail, manufacturing, and financial services will embed these models into the tools you already use. Within 18 months, the forecasting engine sitting behind your demand planning or budgeting software will almost certainly be this type of model, even if the interface looks identical.
The risk for businesses is not missing Toto 2.0 specifically. The risk is arriving at that product evaluation unprepared, with no internal understanding of how these models work, what questions to ask vendors, and what your data actually needs to look like to feed one.
There is also a genuine limitation worth noting. Toto 2.0, like all models in this class, still breaks down when asked to forecast very far into the future, well past the time horizon it was trained on. Old-fashioned statistical methods, the kind that fit a seasonal curve to your data and extrapolate cleanly, still win for long-range projections on simple, stable signals. The best forecasting setups will combine both, not replace one with the other.
Datadog's commercial motivation is transparent: better forecasting inside their monitoring product means fewer false alarms and earlier warnings of system problems for their paying customers. But the open release means the technology leaves their walls. Any reasonably equipped data team at a mid-sized company can start using the same model today that Datadog's own engineers built.