# Google Announces TimesFM-3, A Foundation Model For Multivariate Forecasting > Source: > Published: 2026-09-01 16:21:54+00:00 Google might not be quite at the frontier in general purpose models, but it’s coming up with interesting AI models all the same. Google Research has rolled out TimesFM-3, the latest version of its time-series forecasting model. The headline upgrade is one that businesses running demand planning, finance, or operations forecasts will care about: it can now reason across multiple related time series and external signals at once, instead of just staring at a single series’ own history. The original TimesFM launched back in 2024, and even through last year’s TimesFM-2.5 update, the model family was built strictly for univariate forecasting — predicting a metric using nothing but that metric’s own past values. That’s a bit of a limitation, because most forecasting problems that businesses actually deal with aren’t that clean. A retailer forecasting ice cream sales isn’t just looking at past ice cream sales; sales of related products, foot traffic trends, and upcoming weather or promotions all feed into a good forecast. TimesFM-3 is Google’s attempt to bake that reality into the model itself, rather than leaving it to whoever’s building on top of it. ## What’s Actually New TimesFM-3 comes in at 330 million parameters and was pre-trained on a mix of real-world and synthetic time-series data spanning more than a trillion time points. The core addition is native multivariate support, delivered zero-shot — meaning businesses don’t need to fine-tune the model for their specific use case to get the benefit. Three capabilities stand out: **Multiple targets**— the model can forecast several related series together, such as different product lines, and it outputs both point estimates and quantile ranges for each.**Past covariates**— historical data that isn’t the target itself but plausibly influences it, like past foot traffic, can be fed in as context.**Past-future covariates**— known future events, such as a scheduled promotion or a weather forecast, can be passed in so the model can factor them into what’s ahead, not just what’s already happened. That last category is the more interesting design choice. Google says the model uses a “lookahead” trick for these covariates, where a single token bundles the current data patch together with future patches of the known signal, so the model gets to peek at what’s coming before it commits to a forecast. ## The Architecture Behind It TimesFM-3 keeps the decoder-only transformer backbone of its predecessors, chopping time series into patches of 32 steps before processing them. What changes is how those patches get attended to. The model runs two attention mechanisms in alternation: one that looks across time within a single series (kept strictly causal, so it can’t peek at its own future), and a second that looks across different series at the same point in time, letting the model pick up on cross-series relationships — like how a promotion in one product line might lift or dent sales in another. Google has also dropped the old step-by-step generation approach in favor of producing the entire forecast horizon in one forward pass, using a technique it calls Contiguous Patch Masking. Rather than generating one patch, feeding it back in, and generating the next, the model masks out the future portion of the target series upfront and fills in the whole thing simultaneously — while the known future covariates stay visible throughout. Google frames this as cutting both latency and the compounding errors that come from generating forecasts one step at a time. For each target series, the model outputs nine quantiles per step, so users get a spread of likely outcomes rather than a single number. ## How It Performs Google benchmarked TimesFM-3 against other foundation models on three public forecasting benchmarks — Gift-Eval, FEV-Bench, and Time — measuring both point forecast accuracy and the quality of its probabilistic (quantile) forecasts. The comparison set includes other multivariate-capable models like Chronos-2 and the Toto 2.0 family, alongside TimesFM’s own predecessor, TimesFM-2.5. According to Google, even when TimesFM-3 is run in plain univariate mode — ignoring covariates and cross-series signals entirely — it already ranks ahead of the competing foundation models on average. Switching on full multivariate mode pushes it further ahead across all three benchmarks. ## Availability TimesFM-3 is live now on GitHub and Hugging Face. A BigQuery integration is planned for the coming weeks, following the pattern Google set with TimesFM-2.5, which is already accessible through BigQuery’s `AI.FORECAST` command for users who want to run forecasts without writing ML code. For a company like Google that’s simultaneously racing on frontier language models with Gemini, this kind of steady, less flashy release is a reminder that a lot of its AI research output is aimed squarely at enterprise data problems — demand planning, capacity forecasting, financial projections — rather than chatbots. It’s a pattern that lines up with Google’s broader push to get [AI embedded directly into enterprise data and cloud tooling](https://officechai.com/ai/google-is-hiring-forward-deployed-engineers-to-help-organizations-adopt-ai/), and the breadth of its AI output buttresses its long-running claim to have originated much of the [research underpinning the current AI industry](https://officechai.com/ai/google-came-up-with-90-of-breakthroughs-that-underpin-modern-ai-demis-hassabis/) in the first place.