Google TimesFM 3 Ships Multivariate Forecasting in One Forward Pass
Google Research released TimesFM 3 on 31 August 2026 as a zero shot foundation model for multivariate time series forecasting in a single forward pass, with weights on GitHub and Hugging Face.
What changed
On 31 August 2026, Google Research introduced TimesFM 3, the latest generation of its TimesFM family of zero shot time series foundation models. Research scientists Ayush Jain and Rajat Sen said the model forecasts multiple coevolving time series in one forward pass, trained on a large corpus of time series points, and is now available on GitHub and Hugging Face with BigQuery integration planned in the coming weeks.
Why it matters
Demand planning, energy load, finance, and operations teams still spend weeks stitching univariate models when variables move together. Seasonality shifts, sparse history, and missing entries break pipelines tuned on single series. A multivariate foundation model that runs zero shot could shrink forecast engineering if it holds on real enterprise data rather than public benchmarks alone.
Who is affected
Applied ML leads in retail and logistics, data platform owners on BigQuery, and vendors selling bespoke forecasting services should reassess build versus buy for multivariate workloads before the next planning cycle locks architecture choices.
What to do next
Pull the published weights, benchmark TimesFM 3 against your top three multivariate datasets with the same holdout windows you use for production models, and log where it beats or misses your current stack before committing roadmap spend.
What to watch
Watch for Google BigQuery AI.FORECAST integration landing and any published enterprise case studies with named error metrics on multivariate retail or energy data.
Sources
- Primary. Google Research — TimesFM 3: A zero shot foundation model for multivariate forecasting (31 August 2026). Model scope, release channels, benchmark framing.
- Secondary. Hugging Face — google/timesfm-3.0-pytorch. Public weights channel and license notes.
FAQ
What shipped? <br />TimesFM 3, Google’s zero-shot time-series foundation model with native multivariate forecasting, announced 31 August 2026.
What number should readers lock? <br />About 330 million parameters; pretrained on more than 1 trillion time points per Google Research framing.
What changed vs earlier TimesFM? <br />Earlier checkpoints were univariate; TimesFM 3 is pretrained for multivariate forecasts in one forward pass.
Can companies ship it commercially from public weights? <br />Public weights carry a non-commercial / non-production license caveat in distribution notes — treat production rights as restricted unless Google’s managed path applies.
Where are weights? <br />Google points to GitHub and Hugging Face.
How to read this dispatch
Sources: See Primary/Secondary above (Google Research and Hugging Face linked). <br />What we know vs. what we don’t: Google published the model and benchmark leadership claims among foundation models. Independent replication of every leaderboard is outside this dispatch. <br />Opposing take: Top average ranks on public benches do not automatically equal production rights or domain fit under a non-commercial weight license. <br />Classy aggregates publicly available material; we did not conduct a private interview.